Sliced inference such as SAHI is the standard remedy for small objects in high-resolution imagery, but its gains are usually measured against low-resolution baselines. We revisit sliced inference with scale as the explicit variable and ask when tiling actually pays off. We compare nine dense detectors, eight YOLO models of four generations and FCOS, on VisDrone and DOTA, under pre-specified hypotheses, calibration-only tuning and a single evaluation on held-out test data. Two findings emerge. First, much of the benefit attributed to slicing is a resolution effect: on VisDrone, a single pass trained at a higher resolution outperforms tiling with box fusion (merging decoded boxes) for every detector that could be trained at that resolution, and at equal compute a higher resolution helps more than a larger model. Second, tiles carry complementary information that box fusion wastes. We introduce SCAHI (Slicing with Coverage-Aware Hyper Inference), which accumulates the dense outputs of all passes with weights that give every image position the same total coverage, and reads objects off a peak grid fine enough to keep crowded neighbours apart. On identical passes, SCAHI outperforms box fusion for every YOLO detector on both datasets and matches it for FCOS. On DOTA, whose images far exceed the resolution a single pass can be trained at, tiles with SCAHI beat the best single pass for every detector, by 4.3 to 5.6 mAP50 at 2048 px, and replacing only SAHI's merge step by SCAHI adds 2.3 to 4.0 mAP50. An ablation attributes the gain to the border weighting and the resolution of the peak grid. Source code is available at https://github.com/beckerio/scahi.
Figures & tables
Figure 1: Overview. One detector processes a whole-image pass at its training size and overlapping tiles. The baseline merges decoded boxes across passes (orange). SCAHI accumulates the dense outputs of all passes in a global volume over position and object size, weights them so that every position receives the same total weight, and extracts one peak per object. For YOLOv10 and YOLO26, SCAHI uses the one-to-many head.
Figure 2: Why dense fusion removes tile-border fragments. (a) Box fusion sees two decoded boxes for one object that straddles a tile seam: the fragment predicted by tile A overlaps the full box from tile B too little to be suppressed. (b) Dense fusion weights every prediction by a raised-cosine taper bt over the overlap; the tapers of overlapping tiles sum to one (Eq. 2 ), so a prediction near a tile border carries little weight and the cluster around the object yields one detection. The fusion must also keep adjacent objects apart: its peak search has to be fine relative to object size ( Sec. 5.2 ). Schematic, not to scale.
single pass
1280 px
2048 px + tiles
Δ Q1
Detector
1280
2048
SAHI
+ tiles, box
SCAHI
refined †
val
test
YOLOv10s
33.66
38.47
31.52
35.00
39.34
40.72
+3.47
+2.75
YOLOv10m
36.81
41.03
34.18
37.66
41.29
42.67
+3.37
+3.05
YOLOv11s
32.20
36.80
30.26
34.05
38.66
40.03
+2.75
+1.83
YOLOv11m
35.82
40.00
33.80
37.05
41.54
42.86
+2.95
+1.90
YOLOv12s
32.64
37.94
30.95
34.87
39.80
41.18
+3.07
+2.03
Table 1: VisDrone: single pass vs. tiling (AP on val). Single passes trained and run at 1280 and 2048 px; SAHI; the 1280 px pass with tiles merged by box fusion (NMS); and the 2048 px pass with tiles merged by SCAHI with the pre-specified setting and with the refined grid ( † post hoc on VisDrone). Δ Q1 = single pass at 2048 px minus tiles with box fusion, on val and on the held-out test-dev. YOLOv12m could not be trained at 2048 px. ‡ Added after the verdicts of the other detectors under a dated addendum with unchanged rules.
Figure 3: Resolution or model size (VisDrone val). Single-pass AP against compute for each detector at 1280 and 2048 px (YOLOv10s also at 640, 1536 and 2560 px).
Figure 4: Decomposition on VisDrone test-dev at 2048 px (post hoc). Gain over the single pass with standard post-processing, for SCAHI on the single pass alone, on the single pass plus tiles, and with the refined grid. Tiles add a similar amount for every detector; fusing the single pass alone helps detectors with NMS and costs the NMS-free YOLOv10 and YOLO26; the refined grid makes the overall gain consistent.
Configuration
val
test
GFLOPs
ms
single pass 2048
38.47
33.48
128
9.6
single pass 2560
39.51
33.98
203
13.6
run at 3072 †
39.64
–
298
21.0
2048 + tiles, box fusion
38.85
33.25
562
50.8
2048 + tiles, SCAHI
39.34
34.53
562
77.3
2048 + tiles, SCAHI refined †
40.72
35.28
562
77.3
Table 2: Compute control, YOLOv10s on VisDrone. AP on val and test-dev, compute per image and latency on one RTX 5090, all rows measured in one exclusive run. Box fusion merges on the CPU, SCAHI on the GPU; SCAHI needs the detector’s dense one-to-many outputs and an unfused model, which is included in its latency; its GPU fusion matches the CPU reference up to float32 rounding. Rows marked † are post hoc; the 3072 px row is evaluated on val only.
Configuration
YOLOv10s
YOLOv10m
YOLOv11s
YOLOv12s
FCOS
1024 px
single pass
55.80
58.44
56.16
56.36
42.64
+ tiles, box fusion
64.86
64.57
62.14
62.88
52.56
+ tiles, SCAHI
66.17
67.34
65.26
65.97
52.96
SAHI
59.20
60.96
57.90
60.21
–
SAHI + SCAHI
62.08
63.22
61.87
63.30
–
2048 px
single pass
65.50
67.39
65.19
–
55.01
Table 3: DOTA v1.5 (mAP50 on the held-out val split, devkit, no cap). Tiles at native resolution, merged by box fusion or by SCAHI with the refined grid (pre-specified on DOTA). YOLOv12s could not be trained at 2048 px; its Q3 is measured on its largest trainable pass ( † 1024 px). SAHI + SCAHI: SAHI’s own passes (1024 px weights) merged by SCAHI with the refined grid instead of SAHI’s merge; SAHI rows were not part of the pre-specified FCOS design. Bold: best configuration per detector and input size N . Bottom: pre-specified questions, difference with 95 % CI; verdicts with Holm over YOLOv10s/m and, separately, over YOLOv11s/v12s (both pairs also for Q4); FCOS, a single model, without correction. For YOLOv10, the fusion rows use the one-to-many head; Q1 is pre-specified against box fusion on the one-to-one head (63.18 and 63.97 at 1024 px). FCOS uses a simpler training recipe; compare within a column only.
Variant
VisDrone
DOTA
crowded recall
without normalisation by C
±0.00
±0.00 / ±0.00
±0
without border taper
−0.74 to −1.01
−1.12 / −0.99
−2 to +1
with validity windows
−1.61 to −1.79
– ∗
−1 to −2
refined peak grid
+1.32 to +1.38
+0.36 / +0.41
+18 to +24
Table 4: SCAHI component ablation (post hoc; 2048 px pass with tiles). Difference to the pre-specified setting (border taper, normalisation, cell factor 5) in AP on VisDrone val (range over five detectors) and mAP50 on DOTA-val (YOLOv10s / YOLOv10m), and change in recall of objects with a same-class neighbour.
Result
Dataset
Status
Where (paper or this supplement)
Q1–Q3 verdicts, model extension (five detectors)
VisDrone
pre-specified; val, test-dev once
paper Sec. 5.1–5.2; Sec. 9
Q1–Q3 verdicts, YOLO26s/m (addendum)
VisDrone
pre-specified; val, test-dev once
paper Sec. 5.1–5.2; Sec. 9.1
Criteria of the main study (YOLOv10s)
VisDrone
pre-specified; val, test-dev once
Sec. 10
SCAHI, pre-specified setting ( f=5 , tuned on calib)
both
pre-specified
paper Sec. 4
Refined peak grid ( f=7 – 13 , selected on calib)
VisDrone
post hoc; frozen before test-dev
paper Sec. 5.2
Refined peak grid
DOTA
pre-specified
paper Sec. 5.3
Table 5: Status of the reported results. Pre-specified: questions, rules and settings fixed in writing before the corresponding results existed; post hoc: defined after the relevant results were known (exploratory, never part of a verdict).
Figure 5: Pre-specified questions across detectors (VisDrone model extension). Differences with paired bootstrap 95 % confidence intervals on val (filled) and test-dev (hollow); dashed line: threshold ε=0.30 AP. Q2 is shown for passes at 1280 and 2048 px; the last panel compares the two heads under SCAHI for YOLOv10m and YOLO26. The pooled row covers the five detectors of the original extension; YOLO26s/m, added by addendum, are shown below it. YOLOv12m could only be tested at 1280 px, as it could not be trained at 2048 px. YOLOv10s, the detector of the main study, is reported in Sec. 10 .
Figure 6: Decomposition of the gain of tiles with SCAHI for the detectors of the extension, including YOLO26, on val and test-dev, at 1280 px (top) and 2048 px (bottom): single pass, then SCAHI on the single pass, then adding z=2 tiles, then the refined grid. Error bars: paired bootstrap 95 % confidence intervals of each step. For YOLOv10m and YOLO26, the single pass uses the one-to-one head; YOLOv12m only at 1280 px, as it could not be trained at 2048 px. YOLOv10s is shown in paper Fig. 4.
Figure 7: Recall of crowded and isolated objects (IoU 0.5, val) for the single pass (A), SCAHI on the single pass (B), tiles with SCAHI (C) and the refined grid, for the detectors of the extension; YOLOv12m only at 1280 px, as it could not be trained at 2048 px.
Figure 8: Search of the refined peak grid on calib , detectors of the main study and the original extension. AP against fusion time per image for every searched configuration, per detector, input size N and row (B: single pass, C: with tiles); the selected configuration is marked. The time limit was twice the fusion time of the pre-specified setting.
Figure 9: Search of the refined peak grid on calib , YOLO26s and YOLO26m, with the same grid, time limit and markers as in Fig. 8 .
Detector
single pass
tiles, NMS
tiles, SCAHI
tiles, SCAHI refined
YOLOv10m
7.74 / 15.73
7.48 / 16.39
6.33 / 16.81
6.78 / 17.31
YOLOv11s
5.14 / 12.78
6.33 / 13.92
6.79 / 15.26
6.88 / 15.72
YOLOv11m
5.53 / 14.08
6.24 / 15.28
7.37 / 16.76
7.53 / 17.13
YOLOv12s
4.94 / 13.69
6.09 / 14.65
6.86 / 15.91
7.19 / 16.39
YOLO26s
6.17 / 14.10
6.29 / 14.55
7.10 / 15.10
7.54 / 15.45
YOLO26m
7.76 / 15.93
8.26 / 16.18
7.50 / 16.78
7.76 / 17.21
Table 6: Small objects on test-dev (AP for objects below 8 px and 8–16 px, passes at 2048 px). Single pass with standard post-processing; tiles with box fusion (NMS); tiles with SCAHI, pre-specified and with refined grid. Detectors of the extension that could be trained at 2048 px (YOLOv12m could not); small objects of YOLOv10s are in Tab. 9 .
Figure 10: Compute control for YOLOv10s : AP against GFLOPs and against latency on val (top) and test-dev (bottom), for single passes at 2048 and 2560 px and tiles with SCAHI (pre-specified and refined grid).
Figure 11: Calibration AP during training for all detectors and input sizes; the selected epoch is the maximum on calib.
YOLO26s
YOLO26m
val
test
val
test
Q1
+3.44 [3.03; 3.85]
+2.54
+3.36 [2.91; 3.83]
+2.72
Q2, 1280 px
+1.57 [1.34; 1.83]
+2.36
+1.94 [1.65; 2.21]
+2.24
Q2, 2048 px
+1.63 [1.39; 1.85]
+2.27
+1.77 [1.54; 1.99]
+2.52
Q3
+0.90 [0.64; 1.16]
+1.03
+0.28 [0.04; 0.51]
+0.51
verdicts
Q1, Q2, Q3 met
Q1, Q2 met; Q3 not met
Table 7: YOLO26 addendum (VisDrone; Δ AP with 95 % CI on val, Δ on test-dev; verdict on val).
YOLO26s
YOLO26m
N
configuration
head
val
test-dev
val
test-dev
1280
single pass
one-to-one
33.63
29.94
37.41
32.54
single pass
one-to-many
32.62
28.73
35.99
31.20
SAHI
one-to-one
31.02
28.51
34.56
30.51
+ tiles, box fusion
one-to-one
34.97
30.61
37.89
32.58
+ tiles, box fusion
one-to-many
34.58
30.09
37.09
31.98
Table 8: YOLO26 addendum: all frozen rows (VisDrone, AP on val and test-dev, one evaluation each). N : input size of the single pass; tiles at z=2 . Rows marked † are exploratory and enter no verdict. The questions use: Q1, single pass at 2048 px minus 1280 px + tiles with box fusion (one-to-one head); Q2, SCAHI minus box fusion (one-to-many head); Q3, 2048 px + tiles with SCAHI (one-to-many) minus the single pass at 2048 px (one-to-one).
Figure 20Figure 21
Figure 15: Information loss vs. population. Recall of large objects natively and after degradation by d , compared with originally small objects zoomed to the same apparent size (val, 95 % CIs).
Figure 16: H1 diagnostic : per-size AP of T4 vs. T2 and positives per training chip by zoom level.
box fusion
SCAHI + windows
fusion gain
loss-masked training
34.01
35.33
+1.32
scale-normalised training
34.82
35.92
+1.10
training gain
+0.81
+0.59
Table 12: Training × fusion (val AP, z0.64 + z2 tiles, one-to-many head; SCAHI with validity windows).
Figure 17: False positives near tile borders for box fusion and dense fusion on identical passes (YOLOv10s, val).
Figure 18: Analytic windows : border tapers (sum to 1 over the overlap) and validity windows over log2(σz) (gate G5a: deviation from 1 at most 0.045 ).
Figure 19: Slices of the fused scale-space volume S(x,y,logσ) for four size ranges (blue, darker = higher; orange = ground truth of that size). Small objects light up only in the small- σ slices, larger vehicles in the 18–21 px slice (val image 0000295_02000_d_0000031).
Rows
Δ val
Δ test-dev
AP td
2048 + tiles, box fusion (o2o)
+0.37
−0.23
33.25
2048 + tiles, SCAHI (o2m)
+0.87
+1.05
34.53
1536 + tiles, box fusion (o2o)
+0.92
+0.31
31.92
1536 + tiles, SCAHI (o2m)
+1.86
+1.92
33.52
Table 13: Exploratory rows : single pass + z=2 tiles, Δ AP vs. the single pass of the same weights.
Figure 20: Accuracy vs. cost (val; final latencies with exclusive GPU and CPU). The single-pass front dominates all pre-specified multi-pass rows; only the exploratory 2048 px pass with z=2 tiles and dense fusion lies above it.
Criterion
val
test-dev
(a) selected config. − (1280 + tiles, box) ≥+1.0
+0.92[+0.52;+1.31] not met
+1.81[+1.52;+2.10]
(b) σ<16 : ≥+2.0
+0.42[−0.50;+1.65] not met
−0.78[−1.34;−0.08]
(c) equal AP at ≤0.7× cost
not testable
not testable
(a) vs. single pass 2048
−2.55[−2.98;−2.19] not met
−0.94[−1.25;−0.65]
scale-normalised − loss-masked >ε
−0.01[−0.39;+0.34] not met
+0.29[+0.06;+0.53] not met
SCAHI − box fusion >ε
+1.10[+0.84;+1.37] met
+1.64[+1.44;+1.84] met
Table 14: Main study: pre-specified criteria with confidence intervals (seed 0; paired bootstrap 95 % CI over images, 1,000 resamples; ε=0.30 AP).
Figure 21: Main study: pre-specified criteria on val and test-dev with paired-bootstrap 95 % CIs; dashed lines mark the thresholds.
Figure 22: val vs. test-dev : AP per frozen row, per-size effect and test-dev Pareto view.
YOLOv10s
YOLOv11s
native − single pass 2048 px
−2.14 [ −4.18 ; 0.02 ]
−0.95 [ −2.75 ; 1.28 ]
native − tiles + box fusion
−3.66 [ −6.18 ; −2.43 ]
−1.04 [ −2.59 ; 0.48 ]
native − tiles + SCAHI (refined grid)
−5.85 [ −8.22 ; −4.54 ]
−4.11 [ −5.25 ; −2.43 ]
AP@[.5:.95]
−4.92
−3.47
hybrid − tiles + SCAHI (refined grid), all images
−5.62 [ −7.53 ; −4.49 ]
−4.02 [ −5.12 ; −2.38 ]
hybrid − single pass 2048 px, all images
−0.12 [ −2.86 ; 2.25 ]
+0.59 [ −1.53 ; 3.28 ]
Table 15: Native-resolution single pass on DOTA-val (post hoc; 2048 px weights at z=1 ). Differences in mAP50 (and AP@[.5:.95] where stated) with paired bootstrap 95 % CI. Top: on the images that fit into memory (YOLOv10s 449, YOLOv11s 450 images). Middle: on all images, native where it fits and tiles + SCAHI otherwise. Bottom: native pass minus tiles + SCAHI in AP50 per object size on the images that fit.
Native + 2048 px pass, SCAHI (multi-zoom)
YOLOv10s
YOLOv11s
multi-zoom − tiles + SCAHI
−0.01 [ −1.07 ; 0.54 ]
+0.00 [ −0.80 ; 0.75 ]
AP@[.5:.95]
−0.01
+0.17
AP50, objects >256 px
−0.25
+0.62
multi-zoom − single pass 2048 px
+3.70 [ 2.75 ; 5.09 ]
+3.16 [ 1.84 ; 4.32 ]
multi-zoom − native single pass
+5.85 [ 4.36 ; 7.61 ]
+4.11 [ 2.63 ; 4.94 ]
Table 16: Multi-zoom without tiles and convolutional control on DOTA-val (post hoc). Top: native whole-image pass plus the 2048 px pass, fused by SCAHI (refined grid) or by box fusion, on the images where the native pass fits. Middle: FCOS as a native whole-image pass on all images. Differences in mAP50 with paired bootstrap 95 % CI. Bottom: latency and memory on the images where the native pass fits (YOLOv10s / YOLOv11s).
(a) Equal caps
Contrast
cap 4000
as evaluated
both capped at 200k
without limit
Q2, 1024 px
+0.38 [ −0.94 ; 1.77 ]
+0.40 [ −0.92 ; 1.79 ]
+0.40 [ −0.92 ; 1.79 ]
−1.76
Q2, 2048 px
+0.10 [ −0.96 ; 1.33 ]
+0.09 [ −0.91 ; 1.37 ]
+0.09 [ −0.91 ; 1.36 ]
+0.14
(b) Redundancy: candidates per object, all / crowded / isolated
FCOS
62.7 (44) / 95.4 (85) / 60.1 (41)
YOLOv10s
35.5 (33) / 47.6 (42) / 34.5 (32)
Table 17: FCOS diagnostics and limit check on DOTA-val (post hoc; verdicts unchanged). (a) SCAHI (refined grid) minus box fusion in mAP50 under the pre-specified evaluation (cap 4000), as evaluated, with both fusions capped at the top 200,000 detections per image, and without the pipeline limit. (b) Candidate predictions with IoU ≥0.5 per ground-truth object at 2048 px, summed over all passes, mean (median). (c) Recall at IoU 0.5, SCAHI (refined grid) vs. box fusion. (d) mAP50 of every affected row with and without the 200,000 limit (box fusion: limit per pass).
Detector
N
configuration
mAP50
cap 4000
all objects
AP@[.5:.95]
YOLOv10s
1024
single pass
55.80
55.76
50.59
36.14
tiles + box fusion (o2o head)
63.18
63.15
60.39
41.23
tiles + box fusion
64.86
64.63
61.53
43.50
tiles + SCAHI
66.17
66.02
62.65
44.75
tiles + SCAHI, refined grid
66.17
65.90
62.71
45.63
SCAHI, single pass (no tiles)
56.29
56.30
50.94
36.44
Table 18: All frozen configurations on DOTA-val for the five detectors (devkit mAP50 without cap, with cap 4000, with difficult objects counted, and AP@[.5:.95]). YOLOv10: single pass and SAHI on the one-to-one head, box fusion and SCAHI on the one-to-many head unless marked; YOLOv11s/v12s and FCOS use their single head (FCOS score p⋅centerness ). For box fusion, the pipeline limit applies per pass and the merged output is not capped.
Detector
N
evaluation
single pass
tiles + SCAHI
Δ [95 % CI]
YOLOv10s
2048
mAP50 (primary)
65.50
71.00
+5.50 [+3.55; +7.77]
mAP50, all objects
60.71
67.03
+6.32 [+4.54; +8.64]
mAP50, cap 4000
65.27
71.06
+5.79 [+3.64; +8.18]
AP@[.5:.95]
44.65
50.13
+5.48 [+4.17; +7.03]
YOLOv10m
2048
mAP50 (primary)
67.39
71.73
+4.34 [+3.02; +6.19]
mAP50, all objects
63.06
68.10
+5.04 [+3.63; +6.87]
Table 19: Q3 on DOTA-val under four evaluations : tiles + SCAHI (refined grid) vs. the single pass, both at the largest trainable size, with paired bootstrap 95 % CI. Only the primary metric carries the pre-specified verdict. † YOLOv12s could only be trained at 1024 px, where the single pass is weaker; not comparable with the 2048 px values.
YOLOv10s (2048 px)
YOLOv10m (2048 px)
YOLOv11s (2048 px)
YOLOv12s (1024 px)
FCOS (2048 px)
class
SP
BF
SC
SP
BF
SC
SP
BF
SC
SP
BF
SC
SP
BF
SC
plane
88.6
90.2
90.3
89.9
90.3
90.4
89.3
90.0
90.1
80.6
89.7
89.9
79.7
88.9
89.0
baseball-diamond
70.4
64.8
71.5
72.3
70.1
75.2
71.5
70.9
72.5
66.1
60.9
66.8
60.1
60.5
64.9
bridge
45.5
48.0
51.1
42.7
43.9
47.2
45.2
44.7
48.6
27.7
35.6
39.6
32.6
38.2
41.4
ground-track-field
59.2
68.4
70.6
69.0
67.9
73.3
59.2
62.4
67.7
47.3
58.6
60.6
43.8
45.6
41.7
small-vehicle
70.7
77.7
81.7
75.2
80.5
82.9
72.7
79.1
82.2
58.1
75.3
78.9
57.8
67.9
68.7
Table 20: Per-class AP50 on DOTA-val at the largest trainable input size N : single pass (SP), tiles + box fusion (BF), tiles + SCAHI with refined grid (SC).
YOLOv10s
YOLOv10m
YOLOv11s
YOLOv12s
SAHI
59.20
60.96
57.90
60.21
SAHI passes + SCAHI
61.96
62.98
61.36
63.13
SAHI passes + SCAHI, refined grid
62.08
63.22
61.87
63.30
Δ refined grid (primary, Q4)
+2.87 [ 2.13 ; 4.33 ]
+2.26 [ 1.08 ; 3.25 ]
+3.97 [ 2.93 ; 5.06 ]
+3.10 [ 2.23 ; 4.08 ]
Δ refined grid, cap 4000
+2.80 [ 2.07 ; 4.27 ]
+2.13 [ 0.86 ; 3.16 ]
+4.05 [ 2.91 ; 5.06 ]
+3.02 [ 2.17 ; 4.03 ]
Δ refined grid, all objects
+2.67 [ 1.97 ; 3.98 ]
+1.99 [ 1.11 ; 2.83 ]
+3.92 [ 2.84 ; 4.79 ]
+2.99 [ 2.21 ; 3.95 ]
Table 21: SCAHI on SAHI’s passes (Q4) on DOTA-val (mAP50 without cap unless stated). Differences to SAHI with paired bootstrap 95 % CI; Q4 is decided on the primary metric for the refined grid, Holm over YOLOv10s/m and, separately, over YOLOv11s/12s. Latency (ms per image, fusion on the GPU, optimised path) from one exclusive run; not measured for YOLOv11s/12s.
Detector
single pass
tiles + box fusion
tiles + SCAHI (CPU)
tiles + SCAHI (GPU)
SAHI
YOLOv10s, 2048 px
28.0
175.5
300.9
133.6
194.6
YOLOv10m, 2048 px
34.7
225.2
398.0
186.0
268.1
YOLOv11s, 2048 px
28.5
165.3
286.9
134.2
–
YOLOv12s, 1024 px
21.7
167.1
334.0
142.9
–
Table 22: Latency on DOTA-val (ms per image, mean over 458 images; RTX 5090 at default power, exclusive use, batch 1; all rows from one run on the same machine as Tab. 2 ). Totals including the tile passes. Box fusion and SCAHI on the CPU use the reference passes (tile passes 137.9, 188.7, 125.9 and 133.8 ms); SCAHI on the GPU uses the optimised path of Tab. 2 (passes 120.4, 172.2, 121.0 and 129.6 ms; fusion 13.2–13.8 ms). A one-to-one matching of its detections with the CPU reference shows no deviation beyond float32 rounding. Box fusion runs on the CPU, so the table shows how the cost splits, not which merge step is faster. Peak GPU memory of the SCAHI fusion: 1.6–1.9 GB.
Dataset
Detector
configuration
passes
fusion CPU
fusion GPU
total (CPU)
total (GPU)
VisDrone val
YOLOv10s
single pass 2048
10.3
–
–
10.3
10.3
YOLOv10s
single pass 2560
14.4
–
–
14.4
14.4
YOLOv10s
tiles + SCAHI (pre-specified)
75.6
30.8
9.4
106.4
85.0
YOLOv10s
tiles + SCAHI, refined grid
76.6
60.1
10.2
136.7
86.8
Table 23: Latency breakdown on VisDrone (ms per image, YOLOv10s, exclusive GPU and CPU, RTX 5090 at default power, batch 1, one run): time of the passes and of the fusion on the CPU or the GPU.
Figure 23: When do tiles pay off? Per-image gain of tiles with SCAHI (refined grid) over the single pass at 2048 px, in bins of the share of objects below 16 px and of crowded objects, on VisDrone test-dev and DOTA-val (post hoc; mean with 95 % CI).
Figure 24: Pre-specified questions on DOTA-val : differences in mAP50 with 95 % confidence intervals; dashed: ε .
Figure 25: All frozen configurations on DOTA-val (mAP50, devkit, no cap), for YOLOv10s and YOLOv10m at 1024 and 2048 px: single pass, tiles with box fusion (both heads), SCAHI with the setting from VisDrone and with the refined grid (pre-specified on DOTA), SCAHI on the single pass, and SAHI.
Figure 26: Calibration mAP50 during training on DOTA ; the selected epoch is the maximum on calib.
VisDrone val
Detector
head
configuration
AP
AP50
R50 crowded / isolated
Δ vs. box fusion
Δ vs. refined
YOLOv10s
o2o
box fusion (frozen)
38.85
64.73
79.8 / 91.0
(a) weighted NMS
39.57
64.75
86.2 / 91.5
+0.72 [+0.52; +0.93]
-1.00 [-1.13; -0.84]
(b) weighted WBF
37.56
62.34
87.1 / 93.0
-1.29 [-1.57; -0.98]
-3.01 [-3.25; -2.74]
SCAHI, refined grid
40.57
66.17
85.7 / 92.5
YOLOv10s
o2m
box fusion (post hoc, tuned like the frozen row)
37.85
62.76
79.7 / 87.9
Table 24: Border-weighted box fusion (post hoc, same passes and head). VisDrone val: AP@[.5:.95] / AP50; DOTA-val: mAP50 / AP@[.5:.95]. Recall at IoU 0.5 of crowded / isolated objects. Differences in the first metric with paired bootstrap 95 % CI, against box fusion and against SCAHI with the refined grid; only rows on the same head are compared. Variant (b) was not run for FCOS.
pre-specified setting
refined grid
Dataset
Detector
N
f
bins
τsplat
τpeak
f
bins
τsplat
τpeak
VisDrone
YOLOv10s
2048
5
5
0.005
0.001
11
7
0.00125
0.001
VisDrone
YOLOv10m
1280
5
5
0.005
0.001
13
7
0.00125
0.001
VisDrone
YOLOv10m
2048
5
5
0.005
0.001
13
8
0.00031
0.001
VisDrone
YOLOv11s
1280
5
5
0.005
0.001
11
7
0.00125
0.001
VisDrone
YOLOv11s
2048
5
5
0.005
0.001
11
8
0.00031
0.001
Table 25: Selected SCAHI parameters for tiles with SCAHI (one-to-many head), per dataset, detector and size of the single pass: cell factor f , bins per octave of the size axis, and the thresholds τsplat and τpeak . The pre-specified setting was tuned on calib over thresholds and bins with f=5 ; the refined grid also selects f (post hoc on VisDrone, pre-specified on DOTA). Predictions are clustered at IoU 0.5 in all settings.
Figure 27: Ablation on VisDrone val : difference to the pre-specified SCAHI setting for each variant and detector, with paired bootstrap 95 % confidence intervals, and the sweep of the cell factor. All detectors of the original studies with 2048 px weights; YOLOv12m could not be trained at 2048 px.
Figure 28: Ablation on DOTA-val (post hoc, after the single held-out evaluation). The validity windows are sized for VisDrone and cut off large DOTA objects; they are shown for completeness only.
Figure 29: Qualitative comparison of box fusion (NMS, left) and SCAHI (pre-specified setting, right) on the passes of the 1280 px pass with z=2 tiles (YOLOv10s, val). Green: correct detections; red: false positives; yellow dashed: missed objects (IoU 0.5). Rows: one image per quartile of the per-image AP difference, and the image with the most objects below 16 px.
Figure 30: Qualitative comparison on DOTA-val : box fusion (left) and SCAHI with the refined grid (right), YOLOv10s, 2048 px pass with tiles. Green: correct detections; red: false positives; yellow dashed: missed objects (IoU 0.5).
Deep learning-based object detectors have achieved remarkable success across numerous computer vision applications, yet they continue to struggle with small object detection in high-resolution aerial and satellite imagery, where dense object distributions, variable shooting angles, diminutive target sizes, and substantial inter-class variability pose formidable challenges. Existing slicing strategies that partition high-resolution images into manageable patches have demonstrated promising results for enlarging the effective receptive field of small targets; however, their reliance on fixed slice dimensions introduces significant redundant computation, inflating inference cost and undermining detection speed. In this paper, we propose Adaptive Slicing-Assisted Hyper Inference (ASAHI), a novel slicing framework that shifts the paradigm from prescribing a fixed slice size to adaptively determining the optimal number of slices according to image resolution, thereby substantially mitigating redundant computation while preserving beneficial overlap between adjacent patches. ASAHI integrates three synergistic components: (1)an adaptive resolution-aware slicing algorithm that dynamically generates 6 or 12 overlapping patches based on a learned threshold, (2)a slicing-assisted fine-tuning (SAF) strategy that constructs augmented training data comprising both full-resolution and sliced image patches, and (3)a Cluster-DIoU-NMS (CDN) post-processing module that combines the geometric merging efficiency of Cluster-NMS with the center-distance-aware suppression of DIoU-NMS to achieve robust duplicate elimination in crowded scenes. Extensive experiments on VisDrone2019 and xView, demonstrate that ASAHI achieves state-of-the-art performance with 56.8% on VisDrone2019-DET-val and 22.7% on xView-test, while reducing inference time by 20-25% compared to the baseline SAHI method.
Francesco Moretti, Yi Jin, Guiqin Mario
College of Educational Science and Technology, Polytechnic University of Turin
Slicing-Aided Hyper Inference (SAHI) improves small object detection in high-resolution images but often spends substantial compute on background tiles. We propose region-of-interest (ROI)-Gated SAHI, an inference-time framework that introduces a lightweight proposer to localize foreground regions and restrict sliced refinement to informative areas. We evaluate the framework in two settings. On the COCO128 full split dataset comprising 128 images, static ROI-gating is slower on average than Full SAHI, achieving a speed ratio of 0.88, and yields a lower [email protected] of 0.6602 compared with 0.7569 for Full SAHI. A simple adaptive routing policy with τ= 0.4 educes the mean latency, achieving a slight gain of 1.02× over Full SAHI. On a three-image sparse-to-dense case study, ROI-gating achieves speedups ranging from 0.96× to 6.90× with a mean speedup of 3.41×. These results show that ROI-gating is most beneficial in sparse scenes and requires policy-based routing for robust average behavior.
Rashid Riyadh, Abd Ullah Khan, Imad Gohar +1
Faculty of Information Technology, City University, Malaysia · Department of Information Convergence Engineering, Kyung Hee University, South Korea · Department of Computer Science, National University of Sciences and Technology, Pakistan +3
Tiny object detection remains challenging because tiny instances occupy only a few pixels, and their fine-grained details are easily lost during feature downsampling. Although the highest-resolution feature preserves rich local details for tiny objects, DETR-like detectors often discard it due to its large computational cost, while directly reintroducing this feature also brings dense background noise. To address this problem, we propose ReDHF, a novel framework that Revives the Discarded High-resolution Feature for tiny object detection. ReDHF has three cooperative components. Cross-Scale Deformable Fusion (CSDF) first constructs a semantically enhanced highest-resolution feature and injects its fine-grained details into the shallow feature under a moderate cost. Ellipse-Guided Shallow Supervision (EGSS) then guides the enhanced feature to learn more reliable classification and localization priors by applying geometry-adaptive auxiliary supervision. Full-Scale Query Interaction (FSQI) further enables decoder queries to access both low-level details and high-level semantic information, allowing them to aggregate richer cues for tiny objects. Extensive experiments show that ReDHF consistently improves multiple baselines across different DETR variants and achieves state-of-the-art performance on the evaluated benchmarks. The best ReDHF variants surpass their baselines by +1.4 AP on AI-TOD-v2, +1.6 AP on VisDrone, and +1.5 AP on SODA-D, with especially notable gains on tiny and small objects.
Zehao Wang, Si Wu, Guiping Cao +2
South China University of Technology, Guangzhou, China · Pengcheng Laboratory, Shenzhen, China · Harbin Institute of Technology, Shenzhen, China