Choosing Supervised Post-Processing for Frozen Person Detectors: Information, Object Changes, and Complete Cost
Authors: Tianyi Zhou, Zijun Jia
Organizations: Department of Data Science, City University of Hong Kong, 83 Tat Chee Avenue, Kowloon Tong, Hong Kong SAR, China · School of Automation Science and Electrical Engineering, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing, China
A frozen detector computes hypotheses that can support better final outputs through supervised post-processing. Choosing a readout requires knowing which available information improves detection and what its processing costs. We compare candidate scores, class distributions, geometry, native-set summaries and adapted neighborhood meta classification under common candidate, supervision and operating-point conditions. On 320 held-out CrowdHuman images, adding native-set context to a 97-dimensional readout yields 275 more matches and 96 more unmatched predictions. On a separate 319-image cohort, Full100 matches 6004 objects,versus 5848 for source-selected raw WBF and 6002 for a MetaDetect-style adaptation. Its gain over raw contains 300 added and 144 lost discoveries. After execution-equivalent optimization, Full100 and the meta-statistics readout both take approximately 95 ms per complete warm request on a fixed 64-image workload. Source-frozen transfer gives positive point results on 223 MOT17Det frames from seven sequences, while the main MIAP result shows no improvement. The study supports using sufficiently informative output readouts while retaining optimized meta classification as a competitive option. It also shows why feature dimension and net detection gain alone are unreliable guides to adoption.
Figures & tables
Work
Quality information / action
Relation to this study
MetaDetect [ 22 ]
Class, geometry and neighboring hypotheses; quality estimation
Common-bank validity targets, learner and OP in MetaStats138/141
Detection MetaFusion [ 21 ]
Meta confidence in output filtering
Post-raw-NMS score replacement compared with learned ordering
IoU-Net [ 11 ]
Visual localization confidence; selection and refinement
Different visual/coordinate resources; no original-system run
Learning NMS [ 6 ]
Box relations and matching-dependent suppression
Graph/quality task control has additional supervision and dependencies
Crowd Hat [ 30 ]
Compressed output context; crowd-aware selection
Output reuse precedent; different crowd/counting targets
Table 1: Direct precedents and the interfaces compared here. The executed MetaStats policies adapt published feature families; the other cited systems are methodological precedents rather than numerical reproductions.
Figure 1: Study and deployment interface. (A) Source labels train readouts and select scalar maps and thresholds offline. (B) A request executes one chosen policy with one fixed-detector forward and its required outputs. Full100 and meta-statistics readouts are alternatives. (C) Held-out evaluation compares complete outputs and processing costs.
Source / reference
Endpoint
Difference
Interval
CH / Score2Native3
Recall (pp)
+2.760
[+2.026, +3.496]
CH / Score2Native3
UP/image
-0.060
[-0.323, +0.207]
CH / MetaStats141
Recall (pp)
+0.193
[-0.379, +0.769]
CH / MetaStats141
UP/image
-0.066
[-0.266, +0.129]
MOT17 / Score2Native3
Recall (pp)
+1.269
[+0.507, +2.447]
MOT17 / Score2Native3
UP/image
-0.480
[-1.000, +0.080]
Table 3: Full100 minus each primary reference in the new cohorts. Individual 98.75% intervals target approximate 95% family coverage separately within each source. CH resamples paired images; MOT resamples paired contributions in seven sequence blocks. Negative UP differences indicate fewer unmatched predictions.
Policy
TP
UP
Policy AP50 (%)
Native
5708
1464
74.23
Raw Hard .4
5795
1463
75.53
Raw OP-best WBF
5856
1443
76.53
Score2
5767
1423
75.21
Class91
5708
1424
75.27
ClassGeom97
5719
1408
75.48
Table 4: Info-320 results on 7278 eligible visible-body annotations. The four input groups use their prespecified seed-25, budget- B , Hard-NMS .4 policies. Native and raw anchors are unchanged. AP50 uses the author-visible protocol.
Figure 2: Matched information groups on Info-320. TP and UP are shown relative to Score2. Fitting rows and Hard-NMS are shared, while each policy has its own source-selected threshold.
Figure 3: Objects gained and lost by Full100 relative to strong references. Each row uses actual matched annotation identities; CH-319 and MOT17Det are distinct cohorts.
Figure 4: Complete warm-request means on the fixed 64 images. Original and execution-equivalent optimized Meta138 are compared with Full100 and raw processing. Jobs are repeated engineering measurements.
Policy
TP
UP
Policy AP50 (%)
Raw Gaussian
3520
507
63.68
Raw WBF
3489
486
63.29
Validity WBF
3691
566
67.47
Table 6: Separately fitted RetinaNet readout on CH-255 (5332 objects). Raw Gaussian is the source-selected raw alternative; raw WBF matches the learned output operator. AP50 uses author-visible boxes.
Choice
Evidence-backed implication
Scope
Before learning
Give released-bank raw scores a source OP and operator comparison
Source labels and candidate access required
Simplify the input?
Score5 leaves a quality gap with only a small complete-time saving on the fixed64
No minimal-sufficiency claim
Full100 or Meta?
Retain both; optimized Meta138 is close in quality and time, with different dependencies
No equivalence or universal winner
Add appearance?
ROI/graph points improve some quality coordinates; charge all added dependencies
Stronger-resource alternatives
Change source?
Check intended-source behavior; MOT points and MIAP failure differ
No general transfer rule
Table 7: Adoption choices supported by the evaluated interface and workloads. These recommendations concern the observed policies, not rules predicting success on an arbitrary new source.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Columns
Definition
0–90
logclip(pk,10−7,1) ; background is slot 0, person slot 1.
91–94
(x1/W,y1/H,x2/W,y2/H) .
95
logmax(wnhn,10−12) .
96
logmax(wnW/(hnH),10−9) .
97
maxd∈DiIoU(b,d) .
98
Confidence of the first argmax-IoU box in native output order.
Appendix
Table S8: Full100 features (zero-based columns). Coordinates x1,y1,x2,y2 are in pixels; b denotes their normalization by image width W and height H . Width wn and height hn are floored at 10−9 . Di is the native person set in the same normalized coordinates.
Readout
Readout information
Required dependencies
Raw controls
Original person score
Selected raw operator
Score2Native3
Two log probabilities, three native summaries
Two probability slots; native processing; scalar/head
Full100
91 log probabilities, six geometry terms, three native summaries
All probability slots; native processing; scalar/head
MetaStats138
98 basic outputs and 40 neighborhood statistics
All probability slots; candidate-neighborhood computation; scalar/head
MetaStats141
MetaStats138 plus native summaries
MetaStats138 dependencies plus native processing
Appendix
Table S9: Readout inputs and required processing. Every row starts with the same person-candidate bank. The detector is fixed; learned readouts require source FIT/OP supervision. Native context is computed from native boxes and scores, rather than supplied free of cost. Exact feature definitions are in Supplement S1.
Seed
Input
0.75B
B
1.25B
25
Score2
4226/827/71.39
4422/1084/74.23
4540/1341/75.87
25
Class91
4154/808/70.99
4343/1129/73.70
4455/1430/75.24
25
ClassGeom97
4209/811/71.93
4375/1089/74.33
4485/1388/75.86
25
Full100
4488/845/76.85
4591/1141/78.34
4648/1431/79.13
26
Score2
4198/803/70.95
4422/1088/74.20
4549/1390/75.96
26
Class91
4169/842/71.23
4335/1112/73.62
4448/1435/75.18
Appendix
Table S13: All information-group results on DEV-256. Cells report TP/UP/policy AP50 (%). Budgets are multiples of source-OP Native UP=2388 and apply to OP selection. Seed 25 is primary; seeds 26/27 assess training variation.
Seed
Readout
0.75B
B
1.25B
25
Score2Native3
4274/827/72.33
4455/1099/74.96
4569/1375/76.54
25
MetaStats138
4454/805/76.36
4578/1120/78.17
4645/1412/79.10
25
MetaStats141
4469/815/76.64
4585/1107/78.33
4642/1393/79.12
25
Post-NMS 138
4224/835/72.24
4341/1111/73.92
4447/1396/75.37
25
Post-NMS 141
4334/804/74.39
4445/1077/76.00
4539/1373/77.30
26
Score2Native3
4254/827/72.00
4449/1107/74.82
4552/1397/76.25
Appendix
Table S14: Complete new-reference DEV results. Each cell is TP/UP/policy AP50 (%). Budgets are 0.75B,B,1.25B with B=2388 source-OP UP. Seed 25 is primary; seeds 26/27 describe training variation. These are the same reused 256 development images.
Reference
G
L
Δ TP
Δ UP
Native
405
119
+286
+40
Raw Hard .4
328
129
+199
+41
Raw OP-best WBF
287
149
+138
+61
Score2
347
120
+227
+81
Class91
412
126
+286
+80
ClassGeom97
382
107
+275
+96
Appendix
Table S15: Object-set changes on Info-320, Full100 relative to each reference. G counts annotations matched only by Full100, and L those matched only by the reference. Net changes are Full100 minus reference.
Policy
TP
UP
Ignored
Outputs
Native
5708
1464
1458
8630
Raw Hard .4
5795
1463
1694
8952
Raw OP-best WBF
5856
1443
1797
9096
Score2
5767
1423
1648
8838
Class91
5708
1424
1242
8374
ClassGeom97
5719
1408
1207
8334
Appendix
Table S16: Output accounting on Info-320: outputs = TP + UP + ignored. Ignored predictions are separate from unmatched predictions.
Policy
TP
UP
G
L
AP50 (%)
CrowdHuman (255 images, 5332 eligible objects)
Native
3229
534
74
365
58.35
Raw Gaussian
3520
507
0
0
63.68
Raw WBF
3489
486
137
168
63.29
Validity WBF
3691
566
318
147
67.47
MIAP (256 images, 1119 eligible objects)
Appendix
Table S19: Source-task and zero-tuning transfer results for the separately fitted RetinaNet interface. The adapted policies use Gaussian σ=.5 or WBF .4; G/L are relative to raw Gaussian. AP50 follows the CrowdHuman author-visible or MIAP native-person protocol. Source-fitted heads and thresholds are unchanged on MIAP.
Policy
TP
UP
FN
Outputs
Ignored
AP50 (%)
CrowdHuman, 319 images
Native
5684
1342
1562
8636
1610
74.19
Raw Hard .4
5802
1348
1444
9060
1910
75.92
Raw OP-best WBF
5848
1387
1398
9225
1990
76.77
Full100
6004
1366
1242
9559
2189
80.10
Score2
5782
1319
1464
8943
1842
75.61
Appendix
Table S21: Complete quality and output accounting for the 12 fixed policies in each new cohort. FN counts unmatched eligible annotations; ignored output is separate from UP. AP50 follows the source-specific protocols in S2. All reviewer-call counts are zero.
Job 1
Job 2
Policy
Mean
p95
Mean
p95
Raw Hard .4
89.28
102.12
89.04
101.72
Full100
95.37
108.84
94.77
105.94
Score2Native3
94.22
111.40
92.95
103.02
Meta138 / NumPy
177.64
228.33
176.48
226.08
Meta138 / C99
95.21
108.79
95.04
106.88
Appendix
Table S23: Complete warm requests on the same fixed 64 images (32 CrowdHuman and 32 MOT17Det). Each job averages three real requests within image before aggregation. The two independent jobs reverse policy order; their means and 95th percentiles are reported separately in milliseconds. C99 changes only the execution of frozen MetaStats features. These are engineering workload distributions, not external-population intervals.
Job / source
Policy
Mean
Median
p95
Repeat SD
1 / CH
Raw Hard .4
89.39
82.42
136.07
1.78
Full100
95.31
88.10
140.80
2.04
Score2Native3
94.92
86.95
151.78
4.42
Meta138 / NumPy
177.92
174.58
230.95
3.42
Meta138 / C99
94.88
87.82
141.12
1.73
Meta141 / NumPy
179.88
172.99
235.07
3.94
Appendix
Table S24: Sourcewise complete-request distributions for job 1 of the exact implementation comparison, in milliseconds. Each entry uses 32 image means of three requests. Repeat SD is the mean of within-image sample standard deviations.
Job / source
Policy
Mean
Median
p95
Repeat SD
2 / CH
Raw Hard .4
88.91
81.91
136.33
1.23
Full100
94.92
88.12
144.63
1.52
Score2Native3
92.88
85.99
140.72
1.24
Meta138 / NumPy
176.84
173.01
231.18
4.17
Meta138 / C99
95.48
88.62
146.32
1.75
Meta141 / NumPy
177.76
172.46
231.77
2.50
Appendix
Table S25: Sourcewise complete-request distributions for job 2 of the exact implementation comparison, in milliseconds. Each entry uses 32 image means of three requests. Repeat SD is the mean of within-image sample standard deviations.
Policy
TP
UP
Mean ms
CrowdHuman (64 images)
Raw WBF
871
108
175.58
Validity WBF
890
104
188.45
MIAP (64 images)
Raw WBF
202
87
189.45
Validity WBF
213
104
201.69
Appendix
Table S28: Matched-WBF RetinaNet timing session on 64 images per source, with three interleaved repetitions. TP/UP describe these timed outputs. Policies were unchanged; this session followed disclosure of the quality results.
Cross-city object detection requires a detector trained in one city to generalize to an unlabeled target city. In AI City Challenge 2026 Track 6, we analyze archived configurations of a single RF-DETR-Large detector inside an air-gapped Training-as-a-Service platform whose server returns only an aggregate COCO-style AP over a hidden mixture of source- and target-city images. Frozen 1120 x 1120 inference of a checkpoint trained at 704 x 704 achieved the highest aggregate AP among the evaluated configurations (0.3272 -> 0.3654, +0.0382) without any parameter update, with the largest relative gain on small objects and the largest absolute gain on medium objects, at 2.53x the input pixels. A warm-start 1120px fine-tuning recipe reached 0.3470 while its in-domain validation AP rose (0.767 -> 0.789), a caution that in-domain validation is an unreliable model-selection signal under aggregate-only cross-city feedback. Because that run's evaluation used a higher confidence threshold than the inference-only runs (0.05 vs. 0.01), we treat its score as a descriptive archived outcome rather than a controlled verdict on fine-tuning. Gray-world normalization did not meaningfully change the frozen-1120 result, and a rectangular run was found by audit to have used an unintended portrait orientation. We release verbatim platform commands, configuration snapshots, and an explicit evidence boundary for every claim. Each configuration was submitted once and the best was selected on the hidden server, so these are exploratory, audited findings about the aggregate mixture; they do not establish target-city-specific improvement.
Adapting features before upsampling can improve dense prediction, but the gains may arise from changes in feature values, spatial aggregation, or both. We investigate this ambiguity by comparing matched residual adapters before and after a frozen JAFAR upsampler under a shared parent. We construct a readout-null intervention that preserves a frozen linear classifier's coarse logits while allowing spatial attention weights to adapt. Guidance/value exchanges on trained models and retrained fixed-guidance controls test the utility of this adaptation route. In three-seed experiments on PASCAL Visual Object Classes (VOC) and Cityscapes, adaptation before upsampling improves mean intersection over union (mIoU) by 0.807 and 1.413 percentage points, respectively, over adaptation after upsampling. Across both datasets and the JAFAR and AnyUp upsamplers, readout-null adaptation produces mean gains of 1.379-1.987 points over the corresponding frozen linear baselines. Retrained controls further support the benefit of adaptive spatial guidance. These findings explain how dense predictions can improve despite unchanged coarse classifier outputs and identify spatial guidance as a useful target when adapting frozen feature upsampling systems.
Frozen pretrained image representations are widely used for transfer learning: a backbone is kept fixed, feature vectors are extracted, and a lightweight classifier is trained on top. This pipeline usually feeds the full feature vector to the classifier, even when the target task has far fewer classes than the pretraining task. We revisit a classical alternative: supervised dimensionality reduction with Linear Discriminant Analysis (LDA) before linear probing. We evaluate ten dimensionality-reduction strategies on frozen features from six backbones -- ResNet-18, ResNet-50, MobileNetV3-Small, EfficientNet-B0, ViT-B/16, and DINOv2-ViT-S/14 -- across CIFAR-100, Tiny ImageNet, and CUB-200-2011. Under a fixed logistic-regression protocol, LDA improves accuracy over full features in 11 of 12 coarse-grained configurations, with gains up to 4.5 percentage points while reducing feature dimensionality by 48-87%. The same projection consistently hurts on fine-grained CUB-200, where full features win across all six backbones. This establishes a practical boundary condition: LDA is useful when class-level structure is coarse enough to be captured by mean-separating directions, but it can discard subtle cues needed for fine-grained recognition. We also compare LDA with PCA, PCA+LDA, regularized LDA, Local Fisher Discriminant Analysis, Neighbourhood Components Analysis, and three lightweight LDA extensions. The results show that plain LDA offers the best accuracy-cost tradeoff for most coarse-grained settings, while more complex supervised reduction methods rarely justify their additional cost. Overall, the study provides concrete guidance for when post-hoc supervised projection should, and should not, be inserted into frozen-feature image classification pipelines.