EviScope: Evidence-Conditioned Sampling for Long-Range 3D Detection
Authors: Mingyang Lyu, Yinqian Sun, Gongce Cao, Chujie Tang, Yuhao Zhang, Lu Jia, Qian Zhang, Bo Wang, Pin Lyu
Organizations: Institute of Automation, Chinese Academy of Sciences, Beijing, China · Zhongguancun Academy, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · University of Illinois Urbana-Champaign, USA · University of Science and Technology Beijing, Beijing, China
Query-based camera-radar detectors choose where and over which frames each object query samples from the query's position or feature. We test whether radar evidence around the query should also set these choices at the 150 m range of MAN TruckScenes, where recall differs across object-level radar-return counts within every range band. EviScope is a minimal test: the current frame's returns form a non-learned bird's-eye evidence field, which each query reads at its decoded centre to scale its radial sampling extent and to weight eight frames. It adds 5,450 parameters (0.009%) to RaCFormer, with no latency increase in our timing. Our module-free recipe exceeds the published TruckScenes mAP (25.9 over seven runs against 24.1), so we measure the module only against its own baseline at matched seeds. The evidence a query reads rises with its object's return count (Spearman 0.51) but captures zero-return objects only in part, and the learned radius tracks it (ρ≈−0.91 in four runs). Over seven four-epoch pairs the module changes NDS by +0.31±0.36 and mAP by −0.27±0.22 (mean ± SE). Against a geometry-only conditioner with the same parameter count, trained at three matched seeds, it is higher by +1.25±0.72 NDS and +0.87±0.21 mAP, on mAP in every pair, and the control's mean NDS is near the baseline's. An exploratory far-range gain failed an internally pre-specified held-out test, and large inference-time substitution penalties also occur without query-local evidence: neither learned allocation nor inference-time dependence substitutes for matched end-to-end comparisons.
Figures & tables
Figure 1: (a) Recall of our baseline detector on TruckScenes val by number of radar returns on the object, one curve per range band (Table 20 ). Recall differs across return-count bins within each band, but is not monotonic in return count; zero-return recall is also not monotonic in range. (b) NDS of the same detector with and without EviScope , seven matched seeds, with the paired difference above each pair. The NDS axis starts at 35.
Figure 2: (a) The inherited RaCFormer detector, redrawn compactly; EviScope attaches only to the ray-sampling step of the shared decoder layer. (b) The module. The current frame’s returns form a non-learned field E ; each query reads it at its decoded centre cq ( eq ) and joins its geometry g(q) ; a shared MLP feeds the radial extent αq , the frame weights wq=8πq and a training-only density head (dashed). Sample locations and weights are schematic.
Method
In.
Backbone
Img size
Ep.
NDS ↑
mAP ↑
mATE ↓
mASE ↓
mAOE ↓
mAVE ↓
mAAE ↓
CenterPoint-V
L
Voxel
—
—
35.3
22.6
0.461
0.405
0.468
3.028
0.261
Far3D
C
V2-99
640×960
—
21.4
10.7
0.883
0.507
0.671
1.352
0.338
RCTrans
C+R
R50
256×864
—
22.9
12.6
1.025
0.507
0.608
0.894
0.331
CRT-Fusion
C+R
R50
256×864
—
28.8
16.9
0.833
0.512
0.444
0.852
0.322
BEVFusion
C+R
R50
256×864
—
30.4
18.2
0.941
0.442
0.389
0.892
0.208
SpaRC
C+R
V2-99
640×960
—
35.4
22.5
0.798
0.449
0.476
0.613
0.248
Table 1: TruckScenes val. L : LiDAR, C : camera, R : radar; bold marks the best camera–radar value in a column, the LiDAR row being a reference. Four-epoch rows are seven-run means (Table 7 ), eight-epoch rows one run each; the margin over the published entries is present in the module-free baseline rows. paired Δ is the within-seed difference, not the difference of the rounded means: the seven-pair mean ( ±0.36 NDS, ±0.22 mAP; Section 4.3 ) and the single eight-epoch pair. A V2-99 pair is in Appendix D . ‡ Error terms not retrieved (Appendix C ). § One run at the published ResNet-50 input size, without a paired baseline (Appendix C ).
configuration
epochs
mAP ↑
NDS ↑
mAVE ↓
runs
CBGS, 8 frames (baseline)
4
25.85
38.13
0.754
7
no CBGS, 8 frames †
12
23.10
35.69
0.713
1
CBGS, 4 frames
4
24.02
36.24
0.844
1
CBGS, 2 frames
4
22.36
33.91
0.957
1
Table 2: Recipe ablation on the baseline (no module), ResNet-50. The first row is the seven-run mean of Table 7 ; the others are single runs. † Without class-balanced resampling an epoch has about a quarter the iterations, so this run was given 12 epochs with the same fraction of training before the multi-frame phase; it still sees about 30% fewer samples and differs from the first row in more than one setting.
0–25 m
25–50 m
50–100 m
100–150 m
mAP
NDS
mAP
NDS
mAP
NDS
mAP
NDS
RCTrans
19.9
28.4
12.5
24.1
7.9
19.9
6.2
15.6
CRT-Fusion
26.2
34.4
16.5
27.3
11.0
26.3
7.0
11.4
BEVFusion
28.8
35.7
16.8
25.2
11.7
20.1
7.4
13.9
Horizon3D (R50)
36.5
43.0
23.3
37.8
17.0
32.7
8.9
18.4
Ours (baseline)
41.2
44.5
25.3
39.0
15.7
29.2
8.6
18.1
Table 3: mAP / NDS by disjoint range band, official protocol restricted to each band. Published rows are from Bang et al. (2026) , Table 9; our rows are the seed-0 models. Bold marks the best value in each column.
s0
s1
s2
s3
s4
s5 (held out)
s6 (held out)
Δ recall
+4.96
+0.51
−2.64
+3.27
+2.32
−0.67
−0.59
Table 4: Paired recall difference ( EviScope minus baseline, points) on objects at 100 – 150 m with 3 – 15 returns, the subset where the seed-0 pair gained most (Appendix F ). Seeds 0–4 were examined before the test was written; the internally pre-specified criterion was that both held-out differences be positive.
local evidence
radius
range
radius
window
[0,0.35)
0.734
0–25 m
0.627
5.00
[0.35,0.7)
0.629
25–50 m
0.637
4.94
[0.7,1.1)
0.611
50–100 m
0.663
5.22
≥1.1
0.602
100–150 m
0.692
5.45
>150 m
0.799
5.12
Table 5: Forward-hook probe on the trained EviScope model (seed 0), last decoder layer, 270,000 queries over 300 val samples. radius multiplies the layer’s sampling extent; window is exp(H) of the temporal weights, in frames.
model
substitution
Δ mAP
Δ NDS
Δ recall
EviScope , 4 runs
radial scale → per-frame mean
+0.00±0.11
−0.03±0.18
+0.06±0.07
temporal weights → uniform
−11.4±3.2
−11.0±2.0
−8.2±2.6
EviScope , V2-99
temporal weights → uniform
−13.23
−10.45
−9.81
evidence-blind control
temporal weights → uniform
−8.71
−11.75
—
Table 6: Inference-time substitution: each row replaces one modulation on a trained model and re-evaluates the same weights, with no retraining. The first block is mean ± standard deviation over four ResNet-50 EviScope models; the remaining models are one run each. recall is the fraction of evaluated objects matched within 2 m. The further substitutions and the control model of Section 5.2 are in Appendix P .
Appendix figures & tables28 assets
Supplementary material from the paper’s appendix.
Appendix
s0
s1
s2
s3
s4
s5
s6
mean
mAP
baseline
25.84
26.48
26.17
25.00
25.86
25.82
25.76
25.85
EviScope
26.47
26.51
25.77
24.42
25.79
24.56
25.52
25.58
Δ
+0.63
+0.03
−0.40
−0.58
−0.07
−1.26
−0.24
−0.27
NDS
baseline
38.48
38.85
37.39
37.34
38.60
37.55
38.69
38.13
Appendix
Table 7: The ResNet-50 pair at seven seeds, in the configuration of Section 4.1 ; Δ is EviScope minus baseline within a seed. Seeds 5 and 6 were trained after the held-out test of Section 4.4 was written down.
model
epochs
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAOE ↓
mAVE ↓
mAAE ↓
baseline
4
25.85
38.13
0.655
0.418
0.537
0.747
0.123
baseline
8
26.40
39.24
—
—
—
—
—
EviScope
4
25.58
38.44
0.650
0.418
0.519
0.722
0.126
EviScope
8
26.66
39.91
0.654
0.408
0.480
0.671
0.129
Appendix
Table 8: Four- versus eight-epoch training, ResNet-50 at 256×704 , with the other training settings fixed. Four-epoch rows are seven-run means (Table 7 ); eight-epoch rows are single runs evaluated on different GPU configurations, as described above.
model
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAOE ↓
mAVE ↓
mAAE ↓
baseline
26.7
39.0
0.643
0.410
0.506
0.752
0.124
EviScope
26.6
37.9
0.651
0.410
0.530
0.814
0.135
Appendix
Table 9: The V2-99 pair at 256×704 and four epochs, seed 0, single runs. Every other setting is that of Section 4.1 .
range
0 hits
1–2
3–5
6–15
> 15
0–25 m
+1.5
−0.6
+0.7
+0.3
+1.7
25–50 m
−0.6
+1.5
−1.1
+0.7
+0.3
50–100 m
+1.1
+1.4
+1.2
+2.1
+1.1
100–150 m
+0.8
+0.5
+4.3
+6.2
—
all ranges
+0.6
+0.9
+0.8
+1.4
+0.9
Appendix
Table 10: EviScope minus baseline recall, in points, by range band and radar-hit bin, seed-0 pair only, same protocol as Table 20 . The > 15 cell beyond 100 m holds 23 objects and is omitted. The two bold cells did not replicate on held-out pairs.
configuration
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAOE ↓
mAVE ↓
mAAE ↓
baseline, no module
25.84
38.48
0.663
0.417
0.522
0.711
0.132
+ radius & window
26.46
38.93
0.662
0.418
0.490
0.750
0.110
+ density head ( EviScope )
26.47
40.19
0.636
0.413
0.473
0.665
0.118
control: geometry-only (evidence inputs zeroed)
25.35
38.57
0.675
0.428
0.490
0.703
0.115
control: radius fixed at 0.6× , no module
25.69
38.02
0.660
0.426
0.503
0.780
0.114
control: same, seed 1
25.45
37.63
0.676
0.434
0.549
0.715
0.136
Appendix
Table 11: Component ablation at the full four-epoch schedule, seed 0, single runs. The first three rows are cumulative and share every other setting. The last two rows are the baseline with every query’s sampling extent fixed at 0.6× , the value the trained module most often emits, and no module. The geometry-only control is EviScope with its five evidence inputs set to zero (Appendix T ).
row
radius
window
evid. head
evidence input
mAP
NDS
baseline (no module)
—
—
—
—
14.20
26.65
radius + window
✓
✓
—
full
14.95
27.58
radius only
✓
—
—
full
12.13
25.77
window only
—
✓
—
full
10.46
24.33
full module
✓
✓
✓
full
12.54
22.67
full module
✓
✓
✓
geometry only
12.84
24.74
Appendix
Table 12: Component ablation under a shortened single-epoch schedule. All rows share one recipe (CBGS, one epoch, multi-frame from the first step) and differ only in the module configuration, so the column is internally comparable; it is not comparable to Table 1 , which uses the full four-epoch schedule. radius and window are the two modulation paths; evidence input denotes whether the descriptor sees the evidence field or only its geometric terms.
car
truck
bus
trail.
o.veh.
ped.
moto.
bike
cone
barr.
anim.
sign
Δ AP
+0.7
+1.9
+1.9
+1.3
−4.1
+2.1
−0.3
+1.2
+2.5
−1.3
0.0
+1.7
Appendix
Table 13: Per-class AP change from adding the module ( EviScope minus baseline, same recipe), seed-0 pair only. Per-class differences of this size are within the run-to-run spread of Section 5.3 and we do not read them. Absolute per-class values are in Appendix I .
car
truck
bus
trail.
o.veh.
ped.
moto.
bike
cone
barr.
anim.
sign
baseline (R50)
57.8
37.5
0.1
49.0
11.0
10.4
46.6
12.5
43.3
2.9
0.0
39.0
+ radius & window (R50)
58.0
38.0
0.5
49.4
11.3
10.6
46.8
11.3
43.5
9.5
0.0
38.7
EviScope (R50)
58.5
39.4
2.0
50.3
6.9
12.5
46.3
13.7
45.8
1.6
0.0
40.7
baseline (R50, seed 1)
58.2
39.2
5.3
51.3
9.3
9.4
48.3
12.0
41.0
4.5
0.0
39.4
EviScope (R50, seed 1)
58.7
38.3
0.8
51.5
12.0
7.9
47.3
14.8
40.9
6.9
0.0
39.1
baseline (R50, seed 2)
58.0
37.6
1.5
51.0
8.1
10.2
43.7
10.0
40.5
15.1
0.0
38.5
Appendix
Table 14: Per-class AP (%) on TruckScenes val for the rows we train. Rows share the recipe of Table 1 ; seed 0 unless stated.
parameters
of which module
median latency (ms)
peak memory (GB)
baseline
63,998,040
0
997
3.811
EviScope
64,003,490
5,450
965
3.811
Appendix
Table 17
component
source
Camera branch: ResNet-50, FPN, LSS depth and BEV view transform
Shared decoder layer ( 6 iterations): ray sampling in image, LSS-BEV and radar BEV; per-frame aggregation weights
RaCFormer
Heads, assigner and losses
RaCFormer
Four cameras, 150 m range, 256×256 BEV grid, TruckScenes class mapping
adaptation
Appendix
Table 15: Inherited and added components. RaCFormer : unchanged from Chu et al. (2025) . Adaptation : changes needed to run on TruckScenes, shared by baseline and EviScope ; not a method claim. EviScope : the module of Section 3.2 , the only difference between the paired rows of Table 7 .
0–25 m
25–50 m
50–100 m
100–150 m
car
15.7
32.0
51.8
72.6
truck
3.9
8.1
11.1
23.4
bus
0.2
7.7
12.2
15.8
trailer
2.1
6.4
10.1
23.6
other vehicle
0.3
7.9
7.6
24.9
pedestrian
40.1
70.4
86.2
97.9
Appendix
Table 16: Fraction (%) of annotated objects with no radar return in the current frame, trainval split, against range: by class (top), weather tag (middle) and time-of-day tag (bottom). Cells with fewer than 200 objects are blank; the all classes row is the count-weighted mean.
group
tag
n
mAP
NDS
mATE
all
–
2995
26.5
40.2
0.636
area
highway
1960
28.0
38.1
0.645
city
320
12.4
24.9
0.796
terminal
320
14.3
22.7
0.849
rural
315
16.9
25.2
0.790
weather
clear
1560
24.8
35.8
0.677
Appendix
Table 17: EviScope (R50, seed 0) on val subsets defined by the official scene tags; n is the number of val frames carrying the tag.
0–25 m
25–50 m
50–100 m
100–150 m
group
tag
mAP
NDS
mAP
NDS
mAP
NDS
mAP
NDS
all
–
43.0
47.0
25.3
39.4
16.8
30.5
9.4
19.2
area
highway
42.5
43.5
35.1
43.2
20.6
31.5
10.6
19.1
city
24.8
29.7
6.7
14.5
5.0
13.9
2.7
8.4
terminal
27.0
30.5
12.8
20.0
4.3
10.7
2.7
5.7
rural
39.9
38.6
16.6
24.2
7.2
17.6
4.9
12.8
Appendix
Table 18: EviScope (R50, seed 0), official protocol restricted to the val frames carrying each tag and to objects in each disjoint range band; mAP and NDS (%). Tags with at least 100 val frames.
seed 0
seed 2
epoch
1
2
4
1
2
4
mAP
20.03
23.71
26.47
17.81
21.82
25.77
NDS
31.88
36.51
40.19
29.93
34.41
37.73
Appendix
Table 19: EviScope (R50) evaluated at the end of training epochs, seeds 0 and 2.
Figure 3: (a) Training loss, moving average over 40 iterations, epoch boundaries marked. (b) Spearman correlation of the emitted radius scale with local evidence and with range, probed on the EviScope model at intermediate iterations of the run.
range
0 hits
1–2
3–5
6–15
> 15
0–25 m
82.0 (3,885)
89.3 (4,042)
93.4 (3,455)
96.1 (4,451)
90.2 (1,291)
25–50 m
70.1 (9,950)
83.3 (6,512)
89.1 (3,412)
88.1 (2,542)
92.4 (1,341)
50–100 m
68.9 (14,259)
81.6 (8,688)
80.8 (3,435)
78.0 (2,682)
83.8 (741)
100–150 m
72.2 (8,758)
80.9 (4,243)
74.3 (1,643)
70.7 (899)
—
all ranges
71.4 (36,852)
83.3 (23,485)
85.9 (11,945)
87.5 (10,574)
89.6 (3,396)
Appendix
Table 20: Recall (%) of our baseline detector (seed 0) on TruckScenes val by range × radar-hit count, official 2 m matching, no score threshold; object counts in parentheses, cells below 30 objects omitted. Hit counts are the official num_radar_pts annotation. The last row pools the ranges over all 86,252 evaluated objects.
0 hits
1–2
3–5
6–15
> 15
0–25 m
83.5
88.7
94.1
96.4
91.9
25–50 m
69.5
84.7
88.0
88.9
92.7
50–100 m
70.0
83.0
82.0
80.2
84.9
100–150 m
73.0
81.4
78.5
77.0
65.2
0–25 m
3,885
4,042
3,455
4,451
1,291
25–50 m
9,950
6,512
3,412
2,542
1,341
Appendix
Table 21: EviScope recall (%) by range and by radar returns inside the box (top), and the number of val objects in each cell (bottom); same protocol as Table 20 .
run 0
run 2
run 3
run 4
V2-99
substitution
mAP
NDS
rec.
mAP
NDS
rec.
mAP
NDS
rec.
mAP
NDS
rec.
mAP
NDS
rec.
none
26.48
40.18
80.10
25.77
37.76
79.51
24.38
37.39
79.64
25.79
38.11
79.29
26.55
37.92
80.19
radial mean
26.43
40.01
80.15
25.85
37.98
79.66
24.49
37.22
79.69
25.66
38.12
79.28
26.56
37.86
80.37
radial shuffled
26.44
40.06
80.02
25.70
37.75
79.59
24.43
36.92
79.50
25.70
38.04
79.28
26.53
37.74
80.21
radial constant
26.45
40.14
80.06
25.66
37.55
79.54
24.35
37.41
79.51
25.79
38.18
79.42
26.57
37.86
80.29
temporal uniform
19.60
31.93
75.66
13.68
27.11
71.06
12.07
24.91
69.40
11.37
25.70
69.54
13.32
27.47
70.38
Appendix
Table 22: Absolute mAP / NDS / recall (%) for every substitution and run.
no density head
evidence-blind control
substitution
mAP
NDS
mAP
NDS
none
26.41
38.91
26.02
39.30
radial mean
—
38.81
—
39.25
temporal uniform
19.47
31.57
17.31
27.55
Appendix
Table 23: Absolute mAP / NDS for the two control models under inference-time substitution, seed 0. no density head : radial and temporal modulation trained without the auxiliary density loss. evidence-blind control : the early module that read the evidence field at raw polar coordinates rather than the decoded query centres (Section 3.2 ).
mAP
relative drop
degradation
EviScope
baseline
Δ
EviScope
baseline
none
26.48
25.83
+0.65
—
—
random drop 25%
26.16
25.46
+0.70
−1.2%
−1.4%
random drop 50%
24.94
24.39
+0.55
−5.8%
−5.6%
random drop 75%
21.11
20.42
+0.69
−20.3%
−20.9%
far-range drop 50%
26.23
25.61
+0.61
−0.9%
−0.8%
Appendix
Table 24: Paired radar degradation, four-epoch models. Both models are scored on the same corrupted inputs: the dropout mask is seeded by sample token, so the comparison is per-point paired rather than per-run. Relative drop is measured against each model’s own clean score, so the two columns are not affected by the 0.65 mAP the module already holds when nothing is degraded. sector failure removes a third of the radar sensors.
Figure 4: Bird’s-eye view; ego at the triangle, forward up, rings at 50 , 100 and 150 m. Grey: radar returns. Black: ground-truth boxes. Discs: last-layer queries at their decoded centres, coloured and sized by the emitted radius scale αq .
Figure 5: (a) Returns read at the matched query’s centre (one-cell dilation, log axis) against the returns annotated in the ground-truth box, by range band; boxes span the interquartile range, the bar is the median, whiskers the 5 th– 95 th percentiles; model seed 0, 8,277 matched boxes; cells with fewer than 20 boxes are omitted. (b) ROC of the same reading for separating zero-return from returning boxes, per band; solid seed 0, dashed seed 2.
range
boxes
coverage
ρ0
ρ1
ρ2
AUC 1
0 – 150 m
9,684
0.855 / 0.840
0.48 / 0.48
0.51 / 0.51
0.46 / 0.46
0.72 / 0.72
0 – 50 m
4,058
0.925 / 0.914
0.45 / 0.45
0.47 / 0.47
0.38 / 0.38
0.68 / 0.67
50 – 100 m
3,458
0.820 / 0.802
0.42 / 0.42
0.46 / 0.46
0.39 / 0.40
0.70 / 0.70
100 – 150 m
2,168
0.779 / 0.763
0.33 / 0.35
0.42 / 0.43
0.40 / 0.41
0.72 / 0.72
returns in box
0
1 – 2
3 – 5
6 – 15
>15
median one-cell reading
15.4
21.9
36.7
51.4
111.2
Appendix
Table 25: Ground-truth return counts against the evidence read at the matched query’s centre, last decoder iteration, 300 val samples, two EviScope models (seed 0 / seed 2). ρ : Spearman correlation between the box’s return count and the log-count at dilation 0 , 1 , 2 cells. AUC: separation of zero-return from returning boxes by the one-cell reading. Coverage: share of boxes with a query centre within 2 m. Middle: median one-cell reading and share of zero centre-cell readings by the box’s return count, seed 0. Bottom: the non-learned field read at the true box centres, built from the keyframe sweep that the annotation counts or from the five sweeps per radar the module uses; zero read is the share of zero-return boxes reading below half a return in the centre cell.
Figure 6: Completed within-seed comparisons, first model minus second. Dots are seeds; hollow dots mark arms evaluated on different nodes ( † in Table 26 ). Diamonds and bars are mean ± SE; n counts completed pairs. The four-epoch baseline and V2-99 comparisons, eight-epoch seed-0 pair and density-head seed-0 comparison are also reported in Tables 7 , 8 and 11 , respectively.
comparison
model
seed 0
seed 1
seed 2
mean Δ NDS
four epochs
EviScope a
26.47 / 40.19
26.51 / 40.24
25.77 / 37.73
baseline a
25.84 / 38.48
26.48 / 38.85
26.17 / 37.39
geometry-only
control
25.35 / 38.57†
25.47 / 37.98†
25.31 / 37.87
+1.25
shuffled evidence
control ‡
25.19 / 37.30†
26.31 / 38.25†
–
+2.44
density head
no density branch
26.46 / 38.93b
25.92 / 38.81
26.03 / 38.42
+0.67
Appendix
Table 26: Full-schedule controls; mAP / NDS (%). Superscripts identify results reproduced from other tables for paired reference ( a Table 7 ; b seed-0 row + radius & window of Table 11 ; c seed-0 pair of Table 8 ). Only completed comparisons are reported; a dash denotes no reported pair. Δ is EviScope minus the row’s model within a seed; means use the available pairs. † Arms evaluated on different nodes (Appendix C ); on the control’s node, EviScope scores 26.50 / 40.19 (seed 0) and 26.59 / 40.29 (seed 1). ‡ EviScope with evidence inputs permuted at random across the queries of each sample in training and evaluation.
condition
frames
seed 0
seed 1
seed 2
mean
area
city
320
+1.38
+3.47
+0.92
+1.92
highway
1,960
+0.11
+2.46
+1.22
+1.26
rural
315
+0.46
+0.89
−1.68
−0.11
terminal
320
−0.13
+0.83
−0.21
+0.16
weather
clear
1,560
+1.82
+2.50
−1.72
+0.87
overcast
915
−1.03
+1.85
+1.22
+0.68
Appendix
Table 27: EviScope minus the geometry-only control by official scene condition, Δ NDS (points) at every seed. Conditions with at least 300 of the 2,995 validation frames; each frame carries one tag per category, so rows of different categories overlap. Each subset is scored with the official metrics restricted to its frames. Seed-0 and seed-1 EviScope predictions are those of Table 26 .
Figure 7: Allocation with and without the evidence input, inside each range band. Probe of Table 5 on EviScope (blue, filled) and the geometry-only control (grey, hollow), seeds 0 – 2 of each. Queries are binned by the evidence read at the query: the first bin holds zero readings, the other eight are quantile bins of the positive readings pooled over all models. Top: mean radius scale αq ; bottom: mean effective window exp(H) of the frame weights, in frames. Bins with fewer than 50 queries are omitted. The control’s radius is 0.600 – 0.607 for every query.
comparison
schedule
runs
what differs
Seven matched pairs (Tab. 7 )
4 ep
7 + 7
module on / off, same seed
Eight epochs (Tab. 8 , 26 )
8 ep
2 + 2
module on / off, same seed
Geometry-only (Tab. 26 )
4 ep
3
field channels zeroed in dq
Shuffled evidence (Tab. 26 )
4 ep
2
field channels permuted across queries of each sample
Density head (Tab. 11 , 26 )
4 ep
3
density branch removed
Higher resolution (Tab. 1 )
8 ep
1
256×864 ; no paired baseline
Appendix
Table 28: Summary of experimental settings for the comparisons in this paper (Table 26 , Table 27 , Figure 7 ). Runs counts completed trained models per arm used in the reported comparisons. Inference-time substitutions and probes re-evaluate trained weights without retraining.
Research and Development, Mercedes-Benz AG, Germany. · Department of Computer Science, University of Freiburg, Germany. · Institute for Intelligent Systems, Esslingen University of Applied Sciences, Germany.