Where Grounding Accuracy Lives on the IoU Curve: Label-Free Inference-Time Boundary Refinement

Bo Ma

cs.CVAug 20, 2026

Where Grounding Accuracy Lives on the IoU Curve: Label-Free Inference-Time Boundary Refinement

Authors: Bo Ma

Organizations: Auckland University of Technology

Abstract

Vision--language models can identify the correct referent while returning an imprecise bounding box. We study whether a frozen direct-answer model can use its own prediction to allocate one additional localized observation without accessing target annotations at inference. Label-free precision refinement (LFPR) routes predicted-small regions to a higher-resolution pass, re-grounds the expression inside a context crop, admits a candidate only under fixed geometric guards, and returns a fixed coordinate-wise midpoint. We report results across three evidence tiers. On 31,921 retrospective Ref-L4 expressions, LFPR raises mAcc $_{0.5:0.95}$ from 72.947% to 76.013% ([email protected] 88.531% $\to$ 89.725%, [email protected] 55.788% $\to$ 61.142%). A frozen transfer to 30,969 RefCOCO/RefCOCO+/RefCOCOg expressions improves every dataset at [email protected], mAcc, and mean IoU (pooled mAcc $+0.645$ , [email protected] $+0.817$ ), while [email protected] is unchanged overall: routing alone gains $+1.162$ points there, but crop, guards, and fusion give back $-1.192$ , offsetting rather than showing no strict-IoU effect. A prospective, image-disjoint Flickr30K Entities evaluation improves every endpoint (mAcc $+0.973$ , [email protected] $+1.022$ ), more strongly under a single-box variant (mAcc $+2.575$ , [email protected] $+3.689$ ). The same operator applied to two released grounding specialists improves every endpoint ([email protected] $+1.569$ / $+6.716$ for EGM-4B/8B) at roughly twice the latency, composing with specialist training rather than replacing it. A genuine unguarded control (guard removed from the same candidates) underperforms the incumbent on every metric, showing the guard is load-bearing. Together, these results show that referent selection and boundary precision are partially separable, with different components moving opposing regions of the IoU curve -- behavior a single threshold cannot reveal.

Figures & tables

Evidence tier	$n$ (expr / img)	$B$ [email protected]	$CRG$ [email protected]	$\Delta$ [email protected] [95% CI]	$\Delta$ mAcc [95% CI]
Ref-L4 (retrospective)	31,921 / 9,467	88.531	89.725	+1.194 [0.944, 1.439]	+3.066 [2.818, 3.311]
RefCOCO family (frozen transfer)	30,969 / 3,982	89.212	90.029	+0.817 [0.639, 1.001]	+0.645 [0.469, 0.827]
Flickr30K (prospective, merged-box)	14,481 / 999	74.864	75.678	+0.815 [0.299, 1.324]	+0.973 [0.648, 1.310]

Table 1: Evidence-tier summary:

CRG

against the frozen pre-resolution incumbent

B

, same rows, image-cluster bootstrap intervals. Ref-L4 is retrospective; RefCOCO family and Flickr30K are frozen/prospective transfer that retuned nothing. Full metric detail is in Table 2 and Appendix B.1 .

Dataset	Rows	Images	[email protected] $\Delta$	[email protected] $\Delta$	[email protected] $\Delta$	mAcc $\Delta$
RefCOCO	10,752	1,500	+0.772 [0.516, 1.037]	+0.456 [0.037, 0.881]	$-0.530$ [ $-1.301$ , 0.234]	+0.343 [0.083, 0.594]
RefCOCO+	10,615	1,500	+0.641 [0.381, 0.908]	+0.895 [0.476, 1.319]	+0.377 [ $-0.360$ , 1.110]	+0.644 [0.385, 0.911]
RefCOCOg	9,602	2,600	+1.062 [0.751, 1.391]	+1.250 [0.804, 1.705]	+0.083 [ $-0.587$ , 0.773]	+0.984 [0.698, 1.281]

Table 2: Per-dataset deployed-policy effects,

CRG-B

(final policy against the original pre-routing incumbent

B

), image-clustered 10,000-iteration bootstrap, Holm-corrected within the 12-comparison (3 datasets

\times

4 metrics) family; bold marks adjusted

p<0.05

. Row-weighted deltas equal the pooled

CRG-B=(+0.817,+0.852,-0.029,+0.645)

reported above, verified by a machine-checked weighted-identity assertion (Appendix B.3 ). Image counts overlap across RefCOCO/RefCOCO+; RefCOCOg draws on a largely disjoint image set.

System	Params.	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU	Invalid
Frozen incumbent	8B	89.212	80.790	60.955	75.925	82.262	2
+ resolution routing	8B	89.903	81.597	62.117	76.786	82.911	2
+ LFPR	8B	90.029	81.643	60.926	76.570	82.814	2
Released specialist	4B	90.723	81.856	61.558	76.722	83.171	115
+ LFPR	4B	91.007	82.311	63.128	77.508	83.592	115
Released specialist	8B	91.311	82.027	53.786	74.705	82.365	0

Table 3: Zero-shot released specialists against the frozen incumbent and LFPR, each also shown with LFPR applied to its own box (every delta significant; CIs in Appendix E.1.1 ). LFPR improves every specialist on every metric including Mean IoU, most sharply the 8B specialist’s [email protected] (

+6.72

). Among the base systems, [email protected] (8B specialist) has worst [email protected]/mAcc; the trade persists after refinement. The bottom row is a negative control: the guard gate removed from

CRG

’s own stored candidates. It underperforms every system above it on every metric. Invalid predictions remain in every denominator.

Figure 1: Accuracy against IoU threshold on the 30,969-row transfer split (top) and the paired difference between the guarded system and each specialist (bottom). Specialists lead at permissive thresholds; LFPR crosses the 8B specialist near $t\approx 0.78$ and widens thereafter, but tracks the incumbent and the 4B specialist at strict thresholds until $t=0.95$ . The unguarded negative control trails every system across the full range, reaching only 30.99% at $t=0.95$ against 35.25% for LFPR, 35.43% for the incumbent, and 32.93%/21.60% for the specialists.

Appendix figures & tables38 assets

Supplementary material from the paper’s appendix.

Appendix

Artifact	Rows	Images	Label access	First recoverable access	Recorded uses and decisions
Ref-L4 published test	31,921	9,467	Yes	2026-08-05 or earlier	Baseline scoring; source/size diagnostics; resolution arm; final LFPR scoring and manuscript framing.
Frozen RefCOCO-family transfer	30,969	3,982	Yes	2026-08-10	Frozen LFPR pooled audit; descriptive dataset slices; no target-label tuning.
Shared crop/refinement holdout	2,000	1,713	Labels present in rows	2026-08-08	Selected from the full-test baseline after excluding named prior-pilot images; reused for crop/fusion-only and full LFPR decisions.
Resolution/prompt pilots	800 each	Not fully exported	Yes	2026-08-07–08	Floor, prompt, and stopping choices; some samples were reused and the rung summaries used different references.
Supervised-adaptation split	3,314	Not fully exported	Yes	2026-08-08	Supervised-adaptation control and stopping decision; relation to all other image pools is not reconstructable from the compact export.
Backbone-scale pilot	380	344	Yes	2026-08-09	Same-family 32B direct-model control; no complete LFPR run.

Appendix

Table 4: Minimum data-governance record recoverable from compact manifests and logs. “Not fully exported” is a provenance gap, not evidence of separation.

Dataset	Arm	Rows	SHA-256 (12 hex)
Ref-L4	$B$	31,921	a31497f9f3cb
Ref-L4	$R$ (diagnostic)	31,921	c4d83a2c7800
Ref-L4	$CRG$	31,921	3bee48fa7c3f
RefCOCO family	$B$	30,969	0b0e6a9b508d
RefCOCO family	$CRG$	30,969	fe4508dff83b
Flickr merged-box	$B$	14,481	24170b518493

Appendix

Table 5: Headline artifact hash manifest. Each row is one frozen per-row prediction file (record_key, gt_box, pred_box, and arm-specific fields); every headline table or figure that reports the corresponding arm’s accuracy or cost derives from exactly this file, not a superseded or differently-scored copy of it.

Evaluation	Evidence tier	$n$ (expr / img)	$\Delta$ [email protected]	$\Delta$ [email protected]	$\Delta$ [email protected]	$\Delta$ mAcc	$\Delta$ Mean IoU
Ref-L4 (full split)	Retrospective development	31,921 / 9,467	+1.194	+3.173	+5.354	+3.066	+1.704
RefCOCO/+/g (pooled)	Frozen cross-dataset transfer	30,969 / 3,982	+0.817	+0.852	$-0.029$	+0.645	+0.552
Flickr30K (merged-box)	Prospective, image-disjoint	14,481 / 999	+0.815	+1.326	+1.022	+0.973	+0.712
Flickr30K (single-box)	Prospective, image-disjoint (secondary)	11,574 / 979	+1.175	+3.145	+3.689	+2.575	–

Appendix

Table 6: Evidence-tier summary of the three headline evaluations, in percentage points against each row’s own matched incumbent. Ref-L4 is retrospective development evidence and should not be read as confirmatory; the RefCOCO-family pooled row is frozen transfer that retuned nothing; Table 9 decomposes this same

CRG-B

effect by dataset; the two Flickr30K rows are the paper’s prospective, image-disjoint confirmation, sharing the same 999/979 source images under two different phrase-to-box protocols (Appendix B.4 ) rather than constituting two independent datasets.

Arm	$n$	[email protected]	[email protected]	[email protected]	mAcc	Objects365 $\Delta$
Crop/fusion reused pilot	2,000	+0.70	–	–	–	–
LFPR reused pilot	2,000	+0.85	+2.65	+2.65	+2.055	+1.30
LFPR full test	31,921	+1.194	+3.173	+5.354	+3.066	+1.720

Appendix

Table 7: Historical treatment-minus-baseline effects in percentage points. The two pilot rows are reused and are not independent confirmation. LFPR pilot one-sided LCBs are +0.45, +1.85, +1.61, +1.69, and +0.68 for [email protected], [email protected], [email protected], mAcc, and Objects365 [email protected], respectively. The full test [email protected] paired image-level 95% interval is [+0.944,+1.439]. LCBs explain historical selection and are not confirmatory.

Metric	Incumbent	LFPR	$\Delta$ (pp)	Two-sided 95% CI	Holm $p$
[email protected]	89.212	90.029	+0.817	[0.639, 1.001]	0.0032
[email protected]	80.790	81.643	+0.852	[0.566, 1.147]	0.0032
[email protected]	60.955	60.925	$-0.029$	[ $-0.554$ , 0.497]	0.9311
[email protected]:0.95	75.925	76.570	+0.645	[0.469, 0.827]	0.0032
Mean IoU	82.262	82.814	+0.552	[0.438, 0.668]	—

Appendix

Table 8: Pooled transfer audit on the RefCOCO/RefCOCO+/RefCOCOg family, after the resolution floor was verified to take effect. Intervals are image-cluster bootstrap intervals from the archived paired report; the Holm value is the multiplicity-adjusted primary-family value. [email protected] is the one metric that does not move: the guards and midpoint return the strict-IoU accuracy that the crop and floor earn (Table 11 ). Mean IoU is outside the reported four-metric Holm family (no Holm-adjusted

p

reported here, though its own bootstrap

p<0.001

) and moves in the same direction as [email protected]/mAcc, unlike the strict-IoU metric – consistent with the deployed policy broadly improving boundary precision even though its strict-threshold effect nets to zero after the crop stage gives back part of routing’s [email protected] gain (Table 11 ). Each arm contains two invalid rows, which remain in the denominator.

Dataset	Rows	Images	[email protected] $\Delta$	[email protected] $\Delta$	[email protected] $\Delta$	mAcc $\Delta$
RefCOCO	10,752	1,500	+0.772 [0.516, 1.037]	+0.456 [0.037, 0.881]	$-0.530$ [ $-1.301$ , 0.234]	+0.343 [0.083, 0.594]
RefCOCO+	10,615	1,500	+0.641 [0.381, 0.908]	+0.895 [0.476, 1.319]	+0.377 [ $-0.360$ , 1.110]	+0.644 [0.385, 0.911]
RefCOCOg	9,602	2,600	+1.062 [0.751, 1.391]	+1.250 [0.804, 1.705]	+0.083 [ $-0.587$ , 0.773]	+0.984 [0.698, 1.281]

Appendix

Table 9: Per-dataset deployed-policy effects,

CRG-B

, image-clustered 10,000-iteration bootstrap, Holm-corrected within the 12-comparison (3 datasets

\times

4 metrics) family. Bold marks Holm-adjusted

p<0.05

. Row-weighted deltas equal the pooled

CRG-B=(+0.817,+0.852,-0.029,+0.645)

in Table 8 to within

10^{-10}

. Image counts overlap across RefCOCO/RefCOCO+ (both built on the same COCO subset with different expression styles); RefCOCOg draws on a largely disjoint image set, which is why the three counts do not sum to the pooled 3,982.

Dataset	Boundary shift	Router-caused	Over-expand	Over-shrink	Strict harm rate †
RefCOCO	$95.76\%$	$3.41\%$	$0.72\%$	$0.11\%$	$6.48\%$
RefCOCO+	$95.96\%$	$3.10\%$	$0.80\%$	$0.15\%$	$5.70\%$
RefCOCOg	$91.66\%$	$7.44\%$	$0.82\%$	$0.08\%$	$5.71\%$

Appendix

Table 10: Failure taxonomy and strict-IoU harm rate computed independently on each RefCOCO-family sub-dataset (not pooled), against the true pre-resolution incumbent

B

. Crop-truncation and referent-switch categories are each

0.00\%

on all three and omitted from the columns. † Fraction of

[.9,1]

-baseline rows that fall to

[.75,.9)

after LFPR (the same transition-matrix cell Figure 8 plots for Ref-L4/Flickr), computed per dataset. Zero new inference; reuses the same per-row predictions as Table 9 .

Arm	[email protected]	[email protected]	[email protected]	mAcc	IoU W/L/T vs incumbent
$B$ incumbent	89.212	80.790	60.955	75.925	—
$R$ routing	89.903	81.597	62.117	76.786	1,444 / 651 / 28,874
$C$ crop	89.448	81.462	62.799	76.723	15,725 / 10,608 / 4,636
$CR^{\dagger}$ unguarded negative control	86.974	77.632	56.382	72.707	11,683 / 16,010 / 3,276
$CRG$ full (LFPR)	90.029	81.643	60.925	76.570	13,925 / 12,357 / 4,687

Appendix

Table 11: Five-arm same-row decomposition on the 30,969-row transfer split, all against the same incumbent with 10,000-iteration image-cluster bootstrap and Holm correction. Every arm is significant at adjusted

p\mathrel{\approx}0.003

or better. Routing changes few rows (2,095 of 30,969) but nearly always improves them.

\dagger

CR

here is a genuine unguarded negative control , recomputed directly from

CRG

’s own stored per-row fields (incumbent box, crop candidate, guard decision) with the guard gate removed and every row unconditionally replaced – it uses the exact same crop candidates

CRG

uses, so the comparison is a true ablation of the guard alone. It underperforms the incumbent on every metric, confirming the guard is load-bearing rather than a trade against a competitive alternative. An archived, independently-generated

CR

run ([email protected]/0.75/0.9/mAcc

=89.919/82.311/64.604/77.727

) exists under the same label; that artifact uses an identical guard/fallback/fusion mechanism to

CRG

(verified: its per-row admission decision reproduces

CRG

’s frozen guard thresholds on 30,969/30,969 rows, 100.00% exact) but different upstream baseline and crop-candidate predictions from a separate execution, so it was never a guard ablation at all. That artifact is retained only as provenance of the mislabeling, not as a result.

Resolution routing minus incumbent
Contrast and metric	$\Delta$ (pp)	95% CI
[email protected]	$+0.691$	$[0.554,0.836]$
[email protected]	$+0.807$	$[0.637,0.984]$
[email protected]	$+1.162$	$[0.927,1.402]$
Crop alone (guarded) minus incumbent
[email protected]	$+0.236$	$[0.125,0.356]$

Appendix

Table 12: Same-row component decomposition, 10,000 image-cluster bootstrap resamples, every contrast against the same incumbent. “Crop alone” keeps the frozen guard active (guard and routing are independent axes; this row isolates the crop-only pipeline path, not a guard ablation). The “unguarded crop+routing” row is a true guard ablation, recomputed directly from

CRG

’s own stored candidate and guard fields with the guard gate removed: it underperforms the incumbent on every metric, not merely on [email protected], so cropping does not buy strict-IoU accuracy for free once the guard stops filtering bad candidates. The guards and midpoint in the full system then return a small further [email protected] concession in exchange for the largest [email protected] gain. All contrasts are significant at Holm-adjusted

p\mathrel{\approx}0.003

or better except the full system at [email protected] (

p=0.93

Arm	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU $\Delta$ vs $B$
$B$ incumbent (pre-resolution)	74.864	59.720	37.753	56.975	—
$R$ routing alone	75.893	61.784	39.749	58.585	+1.013
$C$ guarded crop alone	74.926	59.050	37.200	56.463	$-0.111$
$CR^{\dagger}$ unguarded negative control	74.353	58.484	36.662	55.952	$-0.652$
$CRG$ full (LFPR)	75.678	61.046	38.775	57.948	+0.712

Appendix

Table 13: Same-row

B/R/C/CR^{\dagger}/CRG

decomposition on the Flickr30K merged-box confirmation (14,481 rows, 999 images), reconstructed from the frozen run’s own archived per-row fields with no new model inference. Deltas are against

B

, the pre-resolution incumbent, matching Table 11 ’s convention. 10,000-iteration image-cluster bootstrap.

Arm	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU $\Delta$ vs $B$
$B$ incumbent (pre-resolution)	74.864	59.720	37.753	56.975	—
$R_{\mathrm{ALL}}$ unconditional floor	74.629	61.398	42.469	58.747	+0.692
$CR_{\mathrm{ALL}}$ unconditional floor+crop	74.774	61.149	41.551	58.461	+0.591

Appendix

Table 14: Full-population P2-2 unconditional controls on the Flickr30K merged-box confirmation (14,481 rows, 999 images). Mean-IoU deltas are paired against the pre-resolution incumbent

B

; no new inference was performed for the

B

row.

Ref-L4 Midpoint Replacement Guarded [email protected]: 89.725 [email protected]: 61.142 mAcc: 76.013 Mean IoU: +0.691 [email protected]: 89.399 [email protected]: 63.143 mAcc: 76.166 Mean IoU: +0.712 Unguarded [email protected]: 88.262 [email protected]: 59.350 mAcc: 73.864 Mean IoU:

-0.454

[email protected]: 87.247 [email protected]: 62.059 mAcc: 74.491 Mean IoU:

-0.884

Flickr30K merged-box Midpoint Replacement Guarded [email protected]: 75.678 [email protected]: 38.775 mAcc: 57.948 Mean IoU:

-0.301

[email protected]: 74.926 [email protected]: 37.200 mAcc: 56.463 Mean IoU:

-1.123

Unguarded [email protected]: 75.057 [email protected]: 38.071 mAcc: 57.003 Mean IoU:

-0.726

[email protected]: 74.353 [email protected]: 36.662 mAcc: 55.952 Mean IoU:

-1.665

Appendix

Figure 2: Guard

\times

fusion factorial on the retrospective Ref-L4 split and prospective Flickr30K merged-box confirmation. The guarded-midpoint cell (

CRG

) is the shipped policy in both cases; the unguarded cells lose accuracy on every reported threshold, showing that the guard, not the fusion rule alone, is the load-bearing component.

Cell	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU $\Delta$	IoU W/L/T
Guarded, midpoint ( $CRG$ , shipped)	75.678	61.046	38.775	57.948	$-0.301$	5,024 / 5,471 / 3,986
Guarded, replacement	74.926	59.050	37.200	56.463	$-1.123$	4,430 / 6,075 / 3,976
Unguarded, midpoint	75.057	59.747	38.071	57.003	$-0.726$	5,541 / 6,003 / 2,937
Unguarded, replacement ( $CR^{\dagger}$ )	74.353	58.484	36.662	55.952	$-1.665$	4,876 / 6,686 / 2,919

Appendix

Table 15: Guard

\times

fusion factorial on the Flickr30K merged-box confirmation (14,481 rows, 13,360 with an admitted crop candidate; 10,000-iteration image-cluster bootstrap per cell, no new inference). Mean IoU

\Delta

is against the post-routing incumbent

R

(this factorial’s own baseline), not against

B

as in Table 13 above – the two tables use different reference arms by construction, so their

\Delta

columns are not directly comparable. The same pattern as Ref-L4 holds: the guard, not the fusion rule, is what protects accuracy here – both guarded cells beat both unguarded cells at every threshold, and only the unguarded cells fall below the shipped policy. Unlike Ref-L4, guarded replacement does not trade [email protected] for a higher [email protected] on this dataset (

37.200

versus

CRG

’s

38.775

); on Flickr the guarded-midpoint cell dominates guarded-replacement on every metric, so the trade Ref-L4 exhibits is not a universal property of the replacement rule.

Figure 3: Treatment-minus-incumbent accuracy, $\Delta A(t)=A_{\mathrm{LFPR}}(t)-A_{\mathrm{base}}(t)$ , as a function of the IoU threshold on the Flickr30K merged-box confirmation (14,481 rows, 999 images). The darker band is the pointwise 95% interval; the lighter band is the simultaneous (max- $|T|$ ) 95% band used to license concentration claims without cherry-picking a threshold. Shaded vertical strips mark the grid points where the simultaneous band excludes zero. The gain is positive across nearly the entire threshold range rather than concentrated at one operating point.

Rows	Arm	Calls	Mean s	P95 s	Max alloc. GiB
128	$B$	1.000	0.837	0.900	16.44
128	$R$	1.156	1.299	3.847	17.58
128	$C$	2.000	1.646	1.821	16.44
128	$CR$	2.156	2.160	4.717	17.58
128	$CRG$	2.156	2.244	4.734	17.58
128	$R_{\mathrm{ALL}}$	1.000	3.072	3.225	17.59

Appendix

Table 16: Synchronized cost audit on the Flickr30K merged-box confirmation, all seven arms, same protocol and same frozen 128/512-row prefixes as Table 18 .

CR

and

CRG

share the identical call budget and generation calls (they differ only in which box the already-generated crop candidate is turned into – hard replacement versus midpoint fusion – a CPU-only decision that costs nothing), so their latencies match within timing noise, mirroring the same pattern already reported for the RefCOCO-family audit.

R

’s calls exceed 1.000 by exactly its label-free routing rate on this row prefix (e.g.

0.250

at 512 rows, i.e.

25\%

), consistent with the

29.8\%

population-level routing rate reported in Appendix B.5 – the two figures differ because this is a smaller 512-row sample, not a second measurement of the same population. Peak allocated memory is

16.44

GiB for the two arms that never raise resolution (

B

C

) and

17.58

–

17.60

GiB for every arm that can trigger the resolution floor.

Arm	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU $\Delta$ vs $B$
$B$ incumbent (pre-resolution)	74.219	64.648	42.773	59.648	—
$R$ routing alone	75.586	66.602	44.531	61.387	$+1.175$
$C$ guarded crop alone	74.805	63.086	42.578	59.668	$+0.323$
$CR^{\dagger}$ unguarded negative control	74.805	63.086	41.992	59.609	$-0.082$
$CRG$ full (LFPR)	75.000	66.406	43.555	60.723	$+0.932$
$R_{\mathrm{ALL}}$ unconditional floor	75.586	67.773	48.047	62.617	$+1.828$

Appendix

Table 17: All seven arms scored on the identical frozen 512-row Flickr30K merged-box subset (36 images) used for the synchronized cost audit (Table 16 ), so accuracy and cost are now comparable on the same denominator throughout this section.

B

R

C

CR^{\dagger}

CRG

are restricted from the archived full-population run (Table 13 ) by exact record_key match, not rerun;

R_{\mathrm{ALL}}

CR_{\mathrm{ALL}}

are the same predictions already used for the cost audit. Deltas and the 10,000-iteration image-cluster bootstrap are against this table’s own 512-row

B

, not the 14,481-row

B

used elsewhere in this section.

Figure 4: Accuracy-cost trade-off on the Flickr30K merged-box confirmation. Each point uses the same 14,481-row evaluation (accuracy, Table 13 ) and the same synchronized 512-row cost prefix on the same RTX 3090 (cost, this section); no new inference was run to build this figure. $R_{\mathrm{ALL}}$ / $CR_{\mathrm{ALL}}$ (dashed reference lines) cost $3.42\times$ / $6.95\times$ $B$ ; their own matched-sample accuracy is reported in the surrounding text rather than plotted here, since it uses the smaller 512-row sample rather than the full population the other five points use. On this dataset, $R$ alone (not $CRG$ ) is the accuracy-maximizing point on the frontier – see text.

Figure 5: Treatment-minus-incumbent accuracy, $\Delta A(t)$ , as a function of the IoU threshold, across all five evaluations, each against its own true pre-resolution incumbent $B$ . Shaded vertical strips mark each series’ own simultaneous-band significant intervals. Ref-L4 and Flickr (retrospective and prospective evidence respectively) grow larger at strict thresholds; RefCOCO/RefCOCO+/RefCOCOg (shown individually, not pooled) stay positive through most of the range and turn negative only close to $t=0.85$ – $0.92$ , consistent with the deployed policy’s broad per-dataset improvement and softening strict-IoU effect reported at fixed thresholds.

Arm	Calls	Mean s	P95 s	Max alloc. GiB	Invalid %
$B$	1.000	0.945	1.049	16.49	0.39
$C$	1.996	1.876	2.315	16.49	0.39
$R$	1.025	1.048	1.252	17.58	0.39
$CR$	2.021	1.977	2.347	17.58	0.39
$CRG$	2.021	1.865	1.985	17.58	0.39
$R_{\text{ALL}}$	1.000	3.291	3.510	17.59	0.00

Appendix

Table 18: Canonical policy-matched 512-row synchronized cost audit with the resolution floor verified in effect, including the unconditional

R_{\text{ALL}}

CR_{\text{ALL}}

controls. One untimed warm-up plus three CUDA-synchronized timed repeats per row, zero runtime errors on every arm. This table supersedes the earlier

B

C

figures withdrawn above; treat only these absolute latencies, not the earlier ones, as the paper’s cost claim. Selective routing (

R

CR

CRG

) costs little beyond

B

C

because only the flagged small-object rows pay the high-resolution premium; forcing that premium onto every row (

R_{\text{ALL}}

CR_{\text{ALL}}

) costs

3.10

–

3.14\times

more than its selective counterpart at matched call count, quantifying why selective routing is the deployed policy rather than an always-high-resolution default.

Figure 6: Accuracy-cost trade-off on the RefCOCO-family transfer split. Each point uses the same 30,969-row evaluation (accuracy, Table 11 ) and the same synchronized 512-row cost prefix on the same RTX 3090 (latency, Table 18 ); no new inference was run to build this figure. $CRG$ (LFPR, starred) costs $1.865$ s/row against $B$ ’s $0.945$ s/row, a $1.97\times$ increase – consistent with the paper’s “approximately doubles inference latency” characterization, since every admitted row pays one crop pass and router-flagged rows may pay an additional high-resolution pass. The two unconditional controls this cycle added, $R_{\mathrm{ALL}}$ and $CR_{\mathrm{ALL}}$ (dashed reference lines), cost $3.48\times$ and $6.48\times$ $B$ respectively – $3.10$ – $3.14\times$ more than their selective counterparts $R$ and $CR^{\dagger}$ at matched call count (Table 18 ) – for a policy that no longer distinguishes rows worth the extra pass from rows that are not. Their accuracy is now measured on a matched 26,488-row RefCOCO validation split (Table 19 ): both beat the matched $B$ arm across all reported metrics, while $R_{\mathrm{ALL}}$ is stricter-threshold better than $CR_{\mathrm{ALL}}$ (Acc@0.9/mAcc). Because that accuracy readout is on a validation split and this figure’s selective-arm points remain on the frozen test partitions, this panel is still a cost-vs-cost picture rather than a same-split selective-versus-unconditional accuracy frontier. — Figure 6: Accuracy-cost trade-off on the RefCOCO-family transfer split. Each point uses the same 30,969-row evaluation (accuracy, Table 11 ) and the same synchronized 512-row cost prefix on the same RTX 3090 (latency, Table 18 ); no new inference was run to build this figure. $CRG$ (LFPR, starred) costs $1.865$ s/row against $B$ ’s $0.945$ s/row, a $1.97\times$ increase – consistent with the paper’s “approximately doubles inference latency” characterization, since every admitted row pays one crop pass and router-flagged rows may pay an additional high-resolution pass. The two unconditional controls this cycle added, $R_{\mathrm{ALL}}$ and $CR_{\mathrm{ALL}}$ (dashed reference lines), cost $3.48\times$ and $6.48\times$ $B$ respectively – $3.10$ – $3.14\times$ more than their selective counterparts $R$ and $CR^{\dagger}$ at matched call count (Table 18 ) – for a policy that no longer distinguishes rows worth the extra pass from rows that are not. Their accuracy is now measured on a matched 26,488-row RefCOCO validation split (Table 19 ): both beat the matched $B$ arm across all reported metrics, while $R_{\mathrm{ALL}}$ is stricter-threshold better than $CR_{\mathrm{ALL}}$ ([email protected]/mAcc). Because that accuracy readout is on a validation split and this figure’s selective-arm points remain on the frozen test partitions, this panel is still a cost-vs-cost picture rather than a same-split selective-versus-unconditional accuracy frontier.

Arm	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU
$B$ (matched val baseline)	89.218	81.029	62.753	76.475	82.557
$R_{\mathrm{ALL}}$	90.328	83.581	69.394	79.514	84.327
$CR_{\mathrm{ALL}}$	90.464	83.513	68.065	79.275	84.233

Appendix

Table 19: Major H RefCOCO validation-split unconditional-control readout (26,488 rows; 2,738 images), scored on one matched row set with 10,000-iteration image-cluster bootstrap deltas against

B

R_{\mathrm{ALL}}

gains

+1.102

+2.552

+6.641

+3.039

points at [email protected]/0.75/0.9/mAcc (95% CIs

[+0.686,+1.519]

[+1.913,+3.181]

[+5.706,+7.574]

[+2.574,+3.508]

);

CR_{\mathrm{ALL}}

gains

+1.246

+2.484

+5.312

+2.800

(95% CIs

[+0.820,+1.656]

[+1.849,+3.136]

[+4.394,+6.240]

[+2.347,+3.261]

Rows	System	Calls	Gen. tokens	Mean s	P95 s
128	EGM-4B	1.000	98.5	4.798	6.055
128	+ LFPR	2.000	197.4	9.250	11.395
128	EGM-8B	1.000	98.7	4.553	5.148
128	+ LFPR	2.000	196.9	9.116	11.536
512	EGM-4B	1.000	98.9	4.622	5.503
512	+ LFPR	1.998	197.9	9.080	10.861

Appendix

Table 20: Specialist cost stability across audit sizes. Generated tokens are summed across calls; each row has one synchronized timed repeat after a per-row warm-up.

Source	Rows	[email protected] $\Delta$	One-sided LCB
COCO	19,889	+0.875	+0.693
Objects365	12,032	+1.720	+1.260

Appendix

Table 21: Full-split source-stratified effects. Source rows use the same frozen parser and paired image bootstrap; no interaction interval was archived.

Bucket	Rows (%)	[email protected]	Gap to large
Small	6,665 (20.9%)	78.92	$-13.81$
Medium	16,311 (51.1%)	90.15	$-2.58$
Large	8,945 (28.0%)	92.73	–

Appendix

Table 22: Baseline [email protected] by annotation-derived target size, before any resolution intervention.

Arm	[email protected]	[email protected]	[email protected]	mAcc 0.50:0.95
$B$ (direct answer)	72.933	57.500	36.633	54.703
$+$ LFPR ( $CRG$ )	74.233	60.800	37.800	56.673
$\Delta$ (95% CI)	$+1.300$ $[0.494,\,2.122]$	$+3.300$ $[2.297,\,4.349]$	$+1.167$ $[0.236,\,2.085]$	$+1.970$ $[1.377,\,2.572]$

Appendix

Table 23: Qwen3-VL-4B-Instruct, full guarded-fusion pipeline vs. its own direct-answer baseline, on the same 3,000-row stratified Flickr30K merged-box subset used for the cross-family screen. All four 95% paired image-cluster bootstrap CIs exclude zero.

Checkpoint	Dataset (rows)	[email protected]	Parser success	Incumbent gap
GLM-4.1V-9B-Thinking	Flickr merged-box (4,000)	70.18	83.1%	$-4.69$ pp
Qwen2.5-VL-7B-Instruct	Ref-L4-test (3,000, paired)	81.57	97.0%	$-6.47$ pp

Appendix

Table 24: Cross-family checkpoint screening summary. The two rows use different evaluation datasets (noted in column 2) and are not directly comparable to each other in absolute accuracy; each is compared only to the incumbent on its own dataset. Parser success is the fraction of rows producing a usable bare box.

Route	Main result	Disposition
Released EGM transfer	$-7.45$ points at [email protected]; 95% CI $[-9.56,-5.34]$	Closed as a zero-shot transfer; native EGM recipe still needs a matched-compute replication.
Supervised coarse-to-fine	$-0.543$ points at [email protected]; LCB $-1.264$	Closed before the planned larger run.
Transferred parallel box decoding	$-15.80$ points; LCB $-17.64$	Transfer route closed.
SAM3 projection	Positive strict-IoU effects but no headline gate	Closed as a standalone replacement.
SAM3 midpoint fusion	$+0.10$ points at [email protected]; LCB $-0.20$ on fresh data	Retrospective fusion signal did not replicate.
Full crop replacement	$+0.50$ points at [email protected]; LCB $-0.10$	Crop replacement closed; strict-IoU signal retained.

Appendix

Table 25: Matched controls and dispositions.

Figure 7: The LFPR pipeline. The router decision and the guard checks are both computed from already-observed model outputs and image geometry, never from the ground-truth box.

Metric	Incumbent	LFPR	$\Delta$ (pp)	Two-sided 95% CI
[email protected]	88.531	89.725	+1.194	[0.944, 1.439]
[email protected]	77.416	80.590	+3.173	[2.768, 3.584]
[email protected]	55.788	61.142	+5.354	[4.808, 5.906]
[email protected]:0.95	72.947	76.013	+3.066	[2.818, 3.311]
Mean IoU	80.944	82.648	+1.704	[1.542, 1.868]

Appendix

Table 26: Matched full-split results on 31,921 expressions from 9,467 images. Intervals use 10,000 image-cluster bootstrap resamples and describe this retrospective comparison.

\operatorname{mAcc}_{0.5:0.95}

is the manuscript’s primary summary, designated retrospectively rather than confirmed by an external preregistration; [email protected] is the standard benchmark operating point; [email protected]/0.9 are secondary diagnostics; Mean IoU is reported alongside as a threshold-free summary. All five are exploratory and unadjusted here because the comparison itself is retrospective, not because of their role in the metric hierarchy.

Contrast	[email protected]	[email protected]	[email protected]	mAcc
$R$ - $B$ (routing alone)	$+0.711$	$+1.920$	$+3.255$	$+1.773$
	$[0.498,0.932]$	$[1.607,2.244]$	$[2.862,3.658]$	$[1.560,1.994]$
$CRG$ - $R$ (marginal, after routing)	$+0.482$	$+1.253$	$+2.099$	$+1.293$
	$[0.347,0.626]$	$[0.974,1.535]$	$[1.669,2.540]$	$[1.154,1.431]$
$CRG$ - $B$ (deployed policy)	$+1.194$	$+3.173$	$+5.354$	$+3.066$
	$[0.944,1.439]$	$[2.768,3.584]$	$[4.808,5.906]$	$[2.818,3.311]$

Appendix

Table 27: Ref-L4’s routing-alone and post-routing marginal contrasts, 31,921 rows, 9,467 images, 10,000-iteration image-cluster bootstrap (95% CI below each point estimate).

CRG

B

reproduces Table 26 exactly, as it must since both are computed from the same artifacts. Unlike the RefCOCO family (where

CRG

R

’s [email protected] is negative,

-1.192

pp, and unlike Flickr30K (where crop/guard/fusion adds no accuracy beyond routing at any threshold, § 5.4 ), Ref-L4 is the one evidence tier where crop/guard/fusion has a genuine positive marginal effect beyond routing at every reported threshold, not merely a small [email protected]@0.9 trade:

CRG

R

’s [email protected] is

+2.099

pp, not negative. Routing alone still accounts for the majority of the deployed policy’s [email protected] effect (

+3.255

+5.354

pp, 61%), so the same qualitative pattern found on the other two tiers – routing does most of the strict-IoU work – still holds directionally here, but Ref-L4 is not simply a case of the two stages offsetting each other.

Cell	[email protected]	[email protected]	[email protected]	mAcc	Mean IoU $\Delta$	IoU W/L/T
Guarded, midpoint ( $CRG$ , shipped)	89.725	80.590	61.142	76.013	+1.704	19,386 / 10,260 / 2,275
Guarded, replacement	89.399	79.947	63.143	76.166	+1.725	17,333 / 12,317 / 2,271
Unguarded, midpoint	88.262	77.739	59.350	73.864	+0.559	19,764 / 11,321 / 836
Unguarded, replacement	87.247	78.140	62.059	74.491	+0.129	17,612 / 13,490 / 819

Appendix

Table 28: Guard

\times

fusion factorial on the corrected full split (Appendix D.5 ; 10,000-iteration image-cluster bootstrap per cell). Mean IoU

\Delta

and IoU W/L/T are against the true pre-resolution incumbent

B

, cross-checked against the standalone

B

artifact (Appendix D.6 ); an earlier version of this table computed both columns against the post-routing

R

box while this caption described them as against

B

, which understated every cell’s Mean IoU gain and, for the two unguarded cells, had the wrong sign. The guard, not the fusion rule, is what protects [email protected]: both guarded cells beat both unguarded cells at [email protected], and only the unguarded cells fall below the 88.531% incumbent. Midpoint fusion trades some of the guarded arm’s [email protected] for stability – guarded replacement alone reaches 63.143% at [email protected] (its 95% CI is

[+3.512,+4.700]

pp over the incumbent, wider than

CRG

’s

[+1.662,+2.545]

pp) at the cost of a slightly lower [email protected] and more harmed rows (12,317 versus

CRG

’s 10,260).

CRG

is the historically fixed shipped policy – its parameters were set before this factorial was computed, not selected by it; the other three cells are reported for mechanism decomposition, not as alternate deployment candidates.

Figure 8: IoU transition matrix. Diagonal cells dominate on both datasets (70–95% of rows stay in their baseline bin), consistent with LFPR changing a minority of predictions (Section E.7 ). The two off-diagonal cells of most interest: strict-IoU harm, $[.9,1]\to[.75,.9)$ , affects $4.7\%$ of Ref-L4’s near-perfect rows ( $894/18{,}847$ ) and $6.6\%$ of Flickr’s ( $359/5{,}467$ ) – a real, nontrivial cost, not a rounding artifact, consistent with the guard-versus-Acc@0.9 trade reported throughout this paper. Genuine boundary refinement, $[.5,.75)\to[.9,1]$ , is comparatively rare on both datasets ( $0.4\%$ Ref-L4, $1.3\%$ Flickr); most of LFPR’s gain instead comes from the larger $[.75,.9)\to[.9,1]$ cell ( $25.4\%$ Ref-L4, $14.1\%$ Flickr), i.e. rows that were already close and get pushed over the strict threshold, not rows recovered from a coarse miss. — Figure 8: IoU transition matrix. Diagonal cells dominate on both datasets (70–95% of rows stay in their baseline bin), consistent with LFPR changing a minority of predictions (Section E.7 ). The two off-diagonal cells of most interest: strict-IoU harm, $[.9,1]\to[.75,.9)$ , affects $4.7\%$ of Ref-L4’s near-perfect rows ( $894/18{,}847$ ) and $6.6\%$ of Flickr’s ( $359/5{,}467$ ) – a real, nontrivial cost, not a rounding artifact, consistent with the [email protected] trade reported throughout this paper. Genuine boundary refinement, $[.5,.75)\to[.9,1]$ , is comparatively rare on both datasets ( $0.4\%$ Ref-L4, $1.3\%$ Flickr); most of LFPR’s gain instead comes from the larger $[.75,.9)\to[.9,1]$ cell ( $25.4\%$ Ref-L4, $14.1\%$ Flickr), i.e. rows that were already close and get pushed over the strict threshold, not rows recovered from a coarse miss.

Category	Expression (truncated)	Incumbent IoU	Fused IoU
Largest improvement	The middle white boat docked between two other vessels, distinguished…	0.561	0.891
Largest improvement	The table, adorned with a white tablecloth, has a man in a gray shirt…	0.657	0.958
Largest improvement	The table covered with a white tablecloth, at which the man wearing a…	0.644	0.933
Largest harm	An individual in a blue and black jersey is maneuvering the foremost b…	0.985	0.574
Largest harm	Within the cluttered area, a building instrument, namely a hammer, is…	0.940	0.530
Largest harm	A cluster of thin green onions in a blue container.	0.954	0.558

Appendix

Table 29: Fixed-rule qualitative cases from the corrected full split (Appendix D.5 ). The largest-harm rows share a pattern: the incumbent was already close to correct (IoU

>0.9

), and the crop candidate pulled the box toward a plausible but wrong sub-region (a hand, a nearby object, a neighbour in a cluster) that the guard’s geometric checks did not catch. The largest-improvement rows are the mirror case: a coarse incumbent box that the crop resolves to the correct sub-region within a larger scene. Neither pattern is unique to one expression style or object category in this sample.

System	Calls/row	Gen. tokens	Mean s	P95 s	Rows/s	Peak GiB	Invalid
EGM-4B	1.000	98.9	4.622	5.503	0.216	8.45	0.20%
+ LFPR	1.998	197.9	9.080	10.861	0.110	8.45	0.20%
EGM-8B	1.000	97.8	4.581	5.764	0.218	16.48	0.00%
+ LFPR	2.000	196.5	9.154	10.531	0.109	16.48	0.00%

Appendix

Table 30: Synchronized 512-row specialist cost audit on physical RTX 3090 GPUs. Generated tokens are summed across calls. Peak memory is maximum allocated memory; invalid rows remain in the denominator. Each row has one timed repeat after a per-row warm-up.

System	[email protected] $\Delta$	[email protected] $\Delta$	[email protected] $\Delta$	mAcc $\Delta$
EGM-4B	+0.284	+0.455	+1.569	+0.785
	$[0.121,0.449]$	$[0.147,0.776]$	$[1.044,2.099]$	$[0.626,0.940]$
EGM-8B	+0.378	+1.230	+6.716	+2.401
	$[0.227,0.537]$	$[0.915,1.545]$	$[6.010,7.406]$	$[2.226,2.577]$

Appendix

Table 31: Accuracy delta from applying LFPR’s crop/guard/fusion mechanism to each specialist’s own predicted box, full 30,969-row transfer split, 95% paired image-cluster bootstrap intervals below each point estimate. Every interval excludes zero: LFPR improves both specialists on every reported metric, most sharply EGM-8B’s comparatively weak [email protected] (

+6.716

points). Absolute values appear in Table 3 in the main text.

System	Params.	[email protected]
Qwen3-VL-8B + LFPR (ours)	8B	89.73
GLM-4.5V [ GLM-V Team, 2025 ]	106B (MoE)	89.5
GLM-4.6V [ GLM-V Team, 2025 ]	–	88.9
Qwen3-VL-8B incumbent (ours)	8B	88.53
PLaMo 2.1-8B-VL [ GLM-V Team, 2025 ]	8B	86.8
GLM-4.1V-9B-Thinking [ GLM-V Team, 2025 ]	9B	86.8

Appendix

Table 32: Published Ref-L4 test-split [email protected], collected from one leaderboard table for a shared reference point, against our two directly measured rows. Rows are not evaluated under one identical protocol; see text.

Figure 9: Source-stratified accuracy on the recombined predictions. The COCO–Objects365 gap widens as the IoU threshold tightens (+8.23, +12.18, +13.59 points), an association compatible with several localization and domain-shift explanations.

Source	[email protected]	[email protected]	[email protected]	mAcc
COCO (62.3%)	92.32	83.76	63.87	78.81
Objects365 (37.7%)	84.09	71.58	50.28	67.54
Gap	+8.23	+12.18	+13.59	+11.27

Appendix

Table 33: Source-stratified accuracy on the recombined predictions (post-resolution-diagnostic). The gap widens at stricter thresholds and is present within every size bucket, showing that target-size composition alone does not account for the source association.

Vision-Language Grounding

Vision-Language Grounding links words and phrases to the objects, regions, actions, or moments they describe in images, video, and scenes.

See all 418 papers

Aug 18, 2026

Groundbench: Multi-Resolution Polygon Grounding Exposes the Geometry Gap in Vision-Language Models

Bounding-box scores on RefCOCO-family grounding leave little room to distinguish frontier vision-language systems, yet boxes discard object shape. We introduce GroundingBench, a matched benchmark that re-targets the same 1,500 image-expression-referent triples to exact-N polygons at five vertex budgets. A fixed-denominator harness separately audits filled-region intersection over union (IoU) and legal-polygon completion. The strongest tested configuration reaches 88.2 box IoU and 97.1 accuracy at IoU >= .5 ([email protected]), versus 57.7 and 69.2 for direct polygons; because these headline scores use different references, we also compare direct polygons with predicted boxes rasterised against the same contour target, obtaining 57.7 versus 57.3 when pooled. Performance is non-monotone in N and collapses at the densest budget, where legality failures compound residual geometric error. Qwen's thinking-setting contrast is the largest tested input-preserving configuration difference; under frozen templates, false spatial cues are more damaging than false colour cues, and target preference can remain high while contour tracing is poor. Alternate masks and a continuous-area scorer preserve the principal ordering. GroundingBench therefore measures an operational output-geometry gap spanning localisation, boundary construction, serialisation, and topology, rather than latent boundary perception alone.

Zhonghan Bian, Zhenran Wang, Jinsong Li +1

Equal contribution.

Visual Grounding VLM Evaluation Referring Expression Segmentation

Ask PDF

Oct 26, 2026

LOCUS: Locate Once, Check Until Settled. Draft-and-Verify Visual Grounding with a Box Head and Coordinate-Token Verification

Vision-language models (VLMs) ground a referring expression in one of two ways. They decode the box as coordinate tokens, which is precise but takes one sequential step per token, or they regress it from a hidden state with a small head, which takes one step but loses accuracy at high IoU. The two fail differently: a head usually finds the right object but places its edges loosely, while coordinate tokens place edges precisely. We therefore use the regression head as a speculative drafter for the model's own coordinate tokens. Locus writes the head's box as coordinate tokens and checks them with the same backbone by Jacobi iteration over a shared image prefix. The fixed point is the model's greedy coordinate-token answer, reached in about 5.60 sequential passes on average instead of 18.46 decoding steps; head and verifier are one set of weights, trained jointly. On the 8 RefCOCO, RefCOCO+ and RefCOCOg splits, verification raises Locus-2B's [email protected] from 54.87 for its own head to 59.13, and the one-pass head is ahead at [email protected] of every published model with at most 3B parameters except Florence-2-L. Against Jag, the strongest open regression-head model, at its own size and training budget, Locus ties at [email protected] and, with typical acceptance, gains +4.52 points at [email protected], with paired significance tests. Exact answers come 1.5x faster than vLLM greedy-decoding the same coordinate tokens with the same weights; a pure regression head remains faster still.

Muhammad Talha

Independent Researcher

Vision-Language Grounding Visual Grounding Efficient VLM Inference Speculative Decoding

Ask PDF

Oct 7, 2026

InstanceBench: Diagnosing Referential Reasoning and Target Identity in Referring Expression Segmentation

Referring Expression Segmentation (RES) links natural-language descriptions to pixel-level object masks. Yet standard evaluation provides limited insight into instance-level referential reasoning: it does not systematically distinguish referential logics, test target preservation across valid grounding paths, or separate target-selection from mask-generation errors. We introduce InstanceBench, an instance-centered diagnostic benchmark comprising 6,194 images, 9,264 target instances, and 25,077 human-verified expressions. Each target-centric expression set (TCES) fixes the image and target mask while pairing a minimal expression with a same-target variant that uses another valid cue or grounding path. A compact referential-logic taxonomy spans direct target evidence, same-class selection, relational and compositional grounding, and exclusion, while logic-critical construction suppresses simpler shortcuts. Identity-aware metrics measure target retention and set-level success while separating selection from mask-generation errors. Across 22 native-mask RES checkpoints from 18 model families, the strongest checkpoint reaches 67.1% mIoU but only 59.6% [email protected]. Controlled interventions confirm language sensitivity, while failure decomposition identifies target selection rather than mask decoding as the main bottleneck. On a controlled training subset, matched supervision improves identity-aware performance, showing that the diagnosed capability responds to targeted supervision. Collectively, InstanceBench supports a measure-diagnose-improve cycle: measuring target consistency across grounding paths, localizing failure sources, and evaluating targeted interventions.

Yuchen Li, Shaoyang Zhou, Yiran Wang +5

The University of Sydney · The ATLAS Institute · Shanghai Artificial Intelligence Laboratory

Referring Expression Segmentation Referring Expression Comprehension

Ask PDF

Abstract

Vision-Language Grounding

Explore similar work

Groundbench: Multi-Resolution Polygon Grounding Exposes the Geometry Gap in Vision-Language Models

LOCUS: Locate Once, Check Until Settled. Draft-and-Verify Visual Grounding with a Box Head and Coordinate-Token Verification

InstanceBench: Diagnosing Referential Reasoning and Target Identity in Referring Expression Segmentation