CrowdCC: Jointly Benchmarking Corruption Robustness and Deployment Quantization in Crowd Counting

S Mahmudul Hasan; Lu Peng

cs.CVOct 26, 2026

CrowdCC: Jointly Benchmarking Corruption Robustness and Deployment Quantization in Crowd Counting

Authors: S Mahmudul Hasan, Lu Peng

Organizations: Tulane University

Abstract

Crowd-counting models deployed on edge devices face two shifts: inputs are degraded along the capture and transmission pipeline, while models are quantized to satisfy resource constraints. Yet corruption robustness and deployment quantization are typically evaluated separately. We introduce CrowdCC, a controlled corruption benchmark that jointly evaluates these effects using 23 deployment-motivated corruptions at three severity levels, four crowd-counting datasets, and 26 models using A100 TensorRT engines at FP32, FP16, and post-training INT8 precision, paired with throughput and energy measurements from corresponding Jetson Orin NX configurations. The corruptions span six families, with head annotations adjusted when geometry or visibility changes. Our evaluation shows that capture-chain corruptions consistently produce the largest degradation across the evaluated models, while crowd density affects absolute and relative error differently. Clean accuracy does not fully predict corrupted performance. INT8 alters corruption robustness substantially more than FP16, with no systematic direction, and does not consistently improve latency or energy. INT8 inference efficiency is strongly associated with reformat overhead in the compiled engine, while slowdowns are also observed under limited INT8 coverage. Precision changes corrupted-error rankings and robustness-cost frontiers, showing why model selection must consider corruption robustness and measured deployment cost together. The code is available at https://github.com/Tulane-Computer-Architecture-Systems/CrowdCC-Release.

Figures & tables

Figure 1: Corruption sources along the capture pipeline. Four families act during image formation and delivery; scene clutter and visibility loss act beforehand.

(a) base datasets and the instances they generate
Dataset	Inst.	#test-image	Res. (MP)	count min / max	mean
ShanghaiTech A	SHA-C	182	$0.06$ – $1.02$	66 / 2,256	433.3
ShanghaiTech B	SHB-C	316	$0.79$	9 / 539	123.8
UCF-QNRF	QNRF-C	334	$0.08$ – $36.15$	65 / 4,535	718.9
JHU-Crowd++	JHU-C	1,600	$0.05$ – $36.15$	0 / 9,059	326.2

Table 1: CrowdCC at a glance. (a) Full labelled test splits, 2,432 images. (b) 23 corruptions in six families; alt. counts the 7 geometry-altering members.

Panel of Figure 2 — Figure 2: Corruption severity across six families. One UCF-QNRF image is shown for one corruption from each family in Tab. 1 , across all three calibrated severity levels.

			realized engine
Model	$\|\theta\|$ (M)	Output type	FP16 or FP32 ops.	INT8 $\times$	$E_{8}/E_{16}$
CNN–based
CHF [ 39 ]	21.51	D	–	1.99	0.50
DM-Count [ 45 ]	21.50	D	–	1.98	0.48
GCHF [ 39 ]	21.51	D	–	1.98	0.49
CSRNet [ 24 ]	16.26	D	–	1.97	0.49

Table 2: Evaluated models and INT8 deployment efficiency. Output types D/B/P/C denote density maps, blockwise count maps, localization-oriented outputs, and direct count regressors. The FP16 or FP32 ops. column lists operations retained above INT8; superscript

\mathrm{f}

denotes FP32. Speed and energy ratios are four-dataset medians relative to FP16. Bold indicates INT8 engines that are slower or consume more energy than FP16.

Model	SHA	SHB	QNRF	JHU
CHF	57.2/57.3/ 55.7	7.3/7.3/7.9	84.2/84.1/85.7	60.4/60.4/ 60.3
DM-Count	65.7/65.7/66.0	7.8/7.8/7.9	91.5/91.5/91.8	62.4/62.4/62.4
GCHF	59.2/59.2/ 58.1	7.1/7.1/8.3	81.8/81.8/84.1	60.4/60.3/60.5
CSRNet	67.4/67.4/68.2	11.4/11.4/ 11.1	95.8/95.9/96.7	68.9/68.9/ 68.8
P2PNet	57.8/57.8/ 57.5	8.0/8.0/8.1	101.0/101.1/102.0	66.9/66.9/ 65.8
APGCC	57.9/57.9/ 57.3	8.3/8.3/8.6	110.4/110.4/ 105.5	72.7/72.8/73.4

Table 3: Clean MAE by executed precision. Entries are FP32/FP16/INT8. Lower is better. A bold INT8 value is lower than at least one higher-precision value.

Figure 3: FP32 corruption sensitivity across counting metrics. Mean relative degradation across 26 models, four datasets, and three severities, ordered by MAE. † Geometry-altering corruptions use transformed annotations and kept-annotation error.

	mcMAE (FP32/INT8)				FP32		INT8		FP32 RD % by corruption family
Model	SHA	SHB	QNRF	JHU	RD %	mCE	RD %	mCE	Env	Opt	Sen	Enc	Clu	Vis
CHF	74.8/73.6	17.4/16.8	107.0/107.0	72.2/72.8	28.9	0.821	28.5	0.824	62.6	77.5	78.6	38.2	11.5	35.3
DM-Count	81.0/81.1	16.3/15.9	110.0/109.5	73.5/73.0	21.7	0.840	21.1	0.838	38.2	60.4	68.7	32.7	12.8	22.4
GCHF	73.4/72.5	14.2/14.6	104.9/105.9	74.7/75.1	26.1	0.797	25.3	0.812	39.2	60.4	65.6	33.0	18.3	33.1
CSRNet	87.2/87.9	26.5/26.1	120.8/120.1	86.2/85.8	27.8	1.000	26.8	1.000	66.8	71.6	87.9	38.0	7.8	20.5
P2PNet	80.7/82.1	18.7/18.4	117.7/118.4	77.5/77.2	28.1	0.887	30.1	0.892	49.9	62.7	96.8	35.1	15.3	21.8

Table 4: Corrupted error, degradation, and the corruption family that breaks each model. Left: mcMAE per dataset as FP32/INT8, then median RD and mean mCE for each executed precision. Right: FP32 RD by corruption family — Env(ironment), Opt(ics and motion), Sen(sor and exposure), Enc(oding and transport), Clu(tter), Vis(ibility loss). Lower is better throughout. Bold marks the best model in each left-hand column and each model’s own worst family on the right.

Output representation	$n$	RD (%)	mCE	$\|\Delta_{\mathrm{RD}}\|$
D Density map	15	23.6	0.857	2.55
B Blockwise count map	5	22.9	0.916	2.74
P Localization-oriented	4	25.9	0.913	1.97
C Direct count	2	14.0	0.859	0.75

Table 5: Robustness by output representation. Group medians over the models in each group. Lower is better;

|\Delta_{\mathrm{RD}}|

compares INT8 with FP32 and the other columns are FP32.

Figure 4: Crowd-size dependence of corruption error. Images are stratified by dataset-specific clean-count quartiles; parentheses report the median of the four dataset medians.

Figure 5: Counting and localization degradation differ. Each point denotes one FP32 corruption across four native-point models. Dashed: least-squares fit; labels mark the largest deviations on either side.

Panel of Figure 6 — Figure 6: Precision interaction and INT8 execution. In (a) , the band denotes the FP16 95th-percentile range ( $\pm 0.39$ percentage points). In (b) , values below one indicate slowdown; labels denote 1 MCNN, 2 FFNet-MobileNetV3, 3 FFNet-ConvNeXt-T, and 4 ZIP-Small.

Figure 7: Runtime decomposition. Four-dataset median end-to-end time per image for five models with INT8 slowdown (red) and four faster controls. Hatched bars denote INT8; non-GPU time includes preprocessing, transfer, dispatch, and engine residual.

Figure 8: Robustness–energy Pareto frontier. Median INT8 energy on Jetson vs. four-dataset mCE. Shading denotes an equal-cost band; green labels indicate Pareto-undominated models.

Figure 9: Accuracy, robustness, and deployment cost induce different rankings. Each point denotes a model–precision configuration, plotted by Orin energy per image and mean-corruption MAE. The dashed line marks the Pareto frontier; four representative members are labelled.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Sensor & exposure
Corruption	Mechanism	Parameter	Light	Mod.	High
fixed_pattern_sensor_noise	stuck/hot pixels, column FPN [ 13 ]	column $\sigma$	0.002	0.008	0.012
		stuck_fraction	0.00005	0.0002	0.0003
		row_sigma	0.002	0.008	0.012
		temporal_sigma	0.002	0.008	0.012
gaussian_noise	sensor read noise [ 13 ]	$\sigma$	0.010	0.025	0.050

Appendix

Table A.1 : Corruption specification and severity calibration. Mechanisms and severity-dependent parameters for all 23 corruptions at Light, Moderate, and High severity.

\downarrow

denotes parameters that decrease with severity;

\dagger

denotes corruptions requiring transformed head annotations.

			magnitude per dataset (px)
Corruption	Level	spacing frac.	SHA	SHB	QNRF	JHU
median head spacing (px)			17.23	33.11	34.78	32.99
camera motion blur	L	0.050	0.9	1.7	1.7	1.6
	M	0.150	2.6	5.0	5.2	4.9
	H	0.250	4.3	8.3	8.7	8.2
far-field defocus	L	0.050	0.9	1.7	1.7	1.6

Appendix

Table A.2 : Spacing-relative severity calibration. Severity is expressed as a fraction of median nearest-neighbour head spacing; dataset columns report the corresponding pixel magnitudes.

	point	heads removed (%)
Corruption	rule	min	med	max
field-of-view loss	crop	12.3	25.1	80.3
foreground occlusion	occlusion	5.5	16.2	30.7
object distractors	occlusion	6.3	11.9	21.5
packet-loss concealment	occlusion	5.8	11.9	26.3
crowd-like distractors	occlusion	3.6	10.4	29.8

Appendix

Table A.3 : Annotation loss under target-altering corruptions. Fraction of head annotations removed by each transformation: minimum, median and maximum over all evaluated cells.

Panel of Figure 18 — Figure B.1 : Corruption visualizations across severity levels. The same UCF-QNRF image is shown under all 23 corruptions at Light, Moderate, and High severity. Gaussian noise and fixed-pattern noise use native-resolution centre crops because downsampling suppresses their fine-scale effects.

Figure B.2 : Severity response by corruption family. Mean FP32 relative degradation across the three severity levels, grouped by family. All 23 corruptions increase monotonically in aggregate, with far-field defocus showing the steepest rise ( $17.71\%$ – $181.89\%$ ).

Panel of Figure 20 — Figure B.3 : Synthetic weather calibration. All rain and snow severities fall within the JHU-Crowd++ real-weather inter-decile ranges for the selected appearance statistics.

Panel of Figure 21 — Figure B.4 : Corruption-response redundancy. Spearman correlations between FP32 corruption-response profiles, ordered by average-linkage clustering on $1-\rho$ ; markers denote corruption families. The median correlation is $0.52$ , with one pair above $\rho=0.90$ (lens flare and depth-aware haze, $\rho=0.94$ ) and seven additional pairs in $[0.80,0.90)$ .

CNN-based
	ShanghaiTech A		ShanghaiTech B		UCF-QNRF		JHU-Crowd++
Model	Ours	Paper	Ours	Paper	Ours	Paper	Ours	Paper
FFNet-MobileNetV3	70.68/118.90	64.30/109.30	12.64/23.91	–	107.94/198.67	–	65.14/233.29	–
ZIP-Pico	79.88/125.90	71.18/109.60	9.73/15.35	8.23/12.62	111.62/173.88	96.29/161.82	75.96/285.50	–
ZIP-Nano	70.49/109.17	58.86/94.63	8.50/13.69	7.74/12.14	114.44 / 196.29	86.46/147.64	73.53/275.65	–
MobileCount	86.66 / 139.69	89.40/146.00	9.24/ 14.94	9.00/15.40	114.69 / 189.84	131.10/222.60	137.04/397.79	–

Appendix

Table C.1 : Reproduction against published results. Clean FP32 MAE/RMSE on each full test split compared with published values for 26 models. Green indicates a value lower than published, and red marks a relative difference exceeding the 30 % reproduction threshold. Dashes denote unavailable references.

Model	FP32	FP16	INT8	FP32	INT8	FP32	FP16	INT8	$\Delta_{\rm RD}$
	clean MAE			mcMAE		RD (%)
APGCC	57.9	57.9	57.3	75.4	76.1	+30.4	+30.2	+32.7	+2.3
CHF	57.2	57.3	55.7	74.8	73.6	+30.7	+30.7	+32.2	+1.5
GCHF	59.2	59.2	58.1	73.4	72.5	+24.0	+24.0	+24.8	+0.7
CCST	63.6	63.6	63.2	72.3	72.0	+13.8	+13.7	+13.9	+0.2
CCTrans-Full	57.3	57.4	59.7	70.5	73.2	+23.0	+22.8	+22.6	-0.3

Appendix

Table D.1 : Robustness results, ShanghaiTech A. Clean MAE, mcMAE, and RD by precision. RD uses the corresponding clean baseline;

\Delta_{\rm RD}=\mathrm{RD}_{\rm INT8}-\mathrm{RD}_{\rm FP32}

in percentage points.

Model	FP32	FP16	INT8	FP32	INT8	FP32	FP16	INT8	$\Delta_{\rm RD}$
	clean MAE			mcMAE		RD (%)
APGCC	8.3	8.3	8.6	18.5	18.8	+122.5	+122.9	+118.9	-3.6
CHF	7.3	7.3	7.9	17.4	16.8	+139.4	+139.4	+112.8	-26.6
GCHF	7.1	7.1	8.3	14.2	14.6	+100.2	+100.3	+76.8	-23.4
CCST	8.1	8.1	8.3	13.8	13.7	+69.4	+69.7	+65.4	-4.0
CCTrans-Full	7.5	7.5	8.1	12.5	13.7	+66.0	+66.2	+68.8	+2.7

Appendix

Table D.2 : Robustness results, ShanghaiTech B. Clean MAE, mcMAE, and RD by precision. RD uses the corresponding clean baseline;

\Delta_{\rm RD}=\mathrm{RD}_{\rm INT8}-\mathrm{RD}_{\rm FP32}

in percentage points.

Model	FP32	FP16	INT8	FP32	INT8	FP32	FP16	INT8	$\Delta_{\rm RD}$
	clean MAE			mcMAE		RD (%)
APGCC	110.4	110.4	105.5	136.1	131.0	+23.3	+23.2	+24.1	+0.9
CHF	84.2	84.1	85.7	107.0	107.0	+27.1	+27.1	+24.8	-2.3
GCHF	81.8	81.8	84.1	104.9	105.9	+28.2	+28.1	+25.9	-2.3
CCST	96.3	96.2	96.7	106.5	107.5	+10.6	+10.7	+11.2	+0.6
CCTrans-Full	92.3	92.3	91.2	99.5	101.2	+7.8	+7.8	+10.9	+3.1

Appendix

Table D.3 : Robustness results, UCF-QNRF. Clean MAE, mcMAE, and RD by precision. RD uses the corresponding clean baseline;

\Delta_{\rm RD}=\mathrm{RD}_{\rm INT8}-\mathrm{RD}_{\rm FP32}

in percentage points.

Model	FP32	FP16	INT8	FP32	INT8	FP32	FP16	INT8	$\Delta_{\rm RD}$
	clean MAE			mcMAE		RD (%)
APGCC	72.7	72.8	73.4	80.9	81.2	+11.3	+11.2	+10.7	-0.5
CHF	60.4	60.4	60.3	72.2	72.8	+19.6	+19.7	+20.6	+1.0
GCHF	60.4	60.3	60.5	74.7	75.1	+23.7	+23.6	+24.1	+0.5
CCST	61.4	61.4	61.2	67.5	67.9	+9.9	+9.9	+11.0	+1.1
CCTrans-Full	56.8	56.8	59.1	63.8	67.8	+12.3	+12.3	+14.7	+2.4

Appendix

Table D.4 : Robustness results, JHU-Crowd++. Clean MAE, mcMAE, and RD by precision. RD uses the corresponding clean baseline;

\Delta_{\rm RD}=\mathrm{RD}_{\rm INT8}-\mathrm{RD}_{\rm FP32}

in percentage points.

Corruption	F32	I8	F32	I8	F32	I8	F32	I8
	SHA		SHB		QNRF		JHU
far-field defocus	22	24	320	303	29	26	24	23
highlight clipping	40	42	227	223	49	48	30	30
depth-aware haze	39	37	252	244	28	27	18	16
Gaussian noise	49	51	183	182	35	38	58	59
lens flare	23	22	206	196	26	25	10	9

Appendix

Table D.5 : Corruption difficulty. Mean RD across 26 models, ranked by the FP32 four-dataset average. F32/I8 denote precision; † marks geometry-altering corruptions evaluated on kept ground truth.

Sensor & exposure
Corruption	SHA	SHB	QNRF	JHU
clean (no corruption)	67.4	11.4	95.8	68.9
fixed-pattern noise	102.6	28.5	141.0	103.9
Gaussian noise	99.8	33.4	155.6	117.5
highlight clipping	121.1	38.3	179.5	109.4
low light	93.0	38.0	141.5	101.0

Appendix

Table D.6 : mCE reference errors. Three-severity mean evaluated MAE (using kept-annotation normalization where applicable) of CSRNet at FP32 for each corruption and dataset. Each model’s mCE is normalized by the CSRNet reference at the same dataset and precision; the FP16 and INT8 references are obtained analogously and are omitted for brevity. The clean row reports the uncorrupted MAE.

Output representation	$n$	Environment	Optics & motion	Sensor & exposure	Encoding & transport	Scene clutter	Visibility loss
D Density map	15	38.2	60.4	68.7	38.0	14.9	22.4
B Blockwise count map	5	44.2	56.9	53.6	26.7	19.9	25.5
P Localization-oriented	4	51.0	66.3	86.8	32.0	11.9	21.8
C Direct count	2	27.7	37.3	20.8	18.2	17.4	22.5

Appendix

Table D.7 : FP32 relative degradation by output representation. Median RD by corruption family for each representation group; bold indicates the lowest median. Results are descriptive, and the direct-count group contains two Transcrowd variants.

Data	Q	clean	mean $\Delta$	med.	p90	RD	$\Delta$ nMAE
SHA	S	24.1	6.3	2.0	27.9	26.2	4.4
	M	43.8	7.7	2.9	42.8	17.7	3.2
	D	60.0	13.0	5.0	66.1	21.7	3.4
	VD	125.8	33.4	12.0	142.9	26.5	3.7
SHB	S	2.7	3.8	1.3	12.0	141.1	10.1
	M	5.2	5.5	2.0	19.0	105.0	8.0

Appendix

Table D.8 : FP32 robustness by crowd-size quartile. S/M/D/VD denote sparse, medium, dense, and very dense scenes. Statistics aggregate all models, corruptions, and severities; median and 90th-percentile values summarize the upper-tail behavior.

	clean	$F_{1}$ degradation			FP32 drop
Dataset	$F_{1}$	FP32	FP16	INT8	precision	recall
SHA	0.726	0.034	0.033	0.032	0.030	0.036
SHB	0.826	0.051	0.051	0.050	0.003	0.088
QNRF	0.739	0.027	0.027	0.027	0.014	0.037
JHU	0.685	0.029	0.029	0.029	0.005	0.047

Appendix

Table D.9 : Localization degradation by dataset. Clean

F_{1}

and mean corruption-induced degradation for the four native-point models. FP32 precision and recall changes are reported separately; FP16 and INT8 show similar

F_{1}

degradation.

	$F_{1}$ degradation			FP32 drop
Family	FP32	FP16	INT8	precision	recall
Sensor & exposure	0.054	0.054	0.052	0.015	0.085
Optics & motion	0.034	0.034	0.034	-0.001	0.062
Environment	0.033	0.033	0.033	0.018	0.043
Encoding & transport	0.030	0.030	0.030	0.007	0.050
Scene clutter	0.026	0.026	0.025	0.030	0.021

Appendix

Table D.10 : Localization degradation by corruption family. Mean within-dataset

F_{1}

degradation for the four native-point models. FP32 precision and recall changes show that most families primarily reduce recall.

	FP16	INT8
Corruption	med $\Delta_{\rm RD}$	med $\Delta_{\rm RD}$	med $\|\Delta\|$	p10	p90	largest comparison
Sensor & exposure
Gaussian noise	-0.03	+0.58	5.91	-13.2	+20.4	ZIP-Small / JHU
vignetting	+0.01	-0.55	2.80	-16.1	+4.6	ZIP-Small / SHB
highlight clipping	+0.01	-0.52	3.66	-12.8	+9.9	ZIP-Pico / SHB
low light	-0.01	-0.23	3.67	-14.1	+8.5	GCHF / SHB

Appendix

Table E.1 : Precision sensitivity by corruption.

\Delta_{\rm RD}

across 104 paired model–dataset evaluations.

Panel of Figure 34 — Figure E.1 : Corruption-specific INT8 robustness shifts. Distribution of $\Delta_{\rm RD}$ across 104 model–dataset evaluations for each corruption: dots denote medians, boxes the IQR, and whiskers the 10th–90th percentiles. The shaded band marks the FP16 control range ( $\pm 0.99$ percentage points); colors indicate corruption families.

CNN–based
			FP32		FP16		INT8
Model	Par.	INT8	J/img	img/s	J/img	img/s	J/img	img/s
	(M)	%lyr †
APGCC	18.7	94	2.49	9.6	1.17	18.3	0.66	23.6
CHF	21.5	100	2.36	10.4	1.07	20.7	0.52	29.1
GCHF	21.5	100	2.37	10.4	1.07	20.7	0.52	29.1

Appendix

Table F.1 : Deployment performance on ShanghaiTech A. Energy and throughput for FP32, FP16, and INT8. Red marks INT8 regressions relative to FP16. † INT8 layer coverage is measured from A100-built engines and used only as a graph diagnostic.

CNN–based
			FP32		FP16		INT8
Model	Par.	INT8	J/img	img/s	J/img	img/s	J/img	img/s
	(M)	%lyr †
APGCC	18.7	94	3.35	7.3	1.53	14.1	0.86	18.1
CHF	21.5	100	3.54	7.1	1.59	14.5	0.73	20.9
GCHF	21.5	100	3.54	7.1	1.59	14.6	0.73	20.9

Appendix

Table F.2 : Deployment performance on ShanghaiTech B. Energy and throughput for FP32, FP16, and INT8. Red marks INT8 regressions relative to FP16. † INT8 layer coverage is measured from A100-built engines and used only as a graph diagnostic.

CNN–based
			FP32		FP16		INT8
Model	Par.	INT8	J/img	img/s	J/img	img/s	J/img	img/s
	(M)	%lyr †
APGCC	18.7	94	12.35	1.6	6.46	2.4	4.28	2.8
CHF	21.5	100	14.75	1.5	7.29	2.4	4.19	2.9
GCHF	21.5	100	14.74	1.5	7.23	2.4	4.16	2.9

Appendix

Table F.3 : Deployment performance on UCF-QNRF. Energy and throughput for FP32, FP16, and INT8. Red marks INT8 regressions relative to FP16. † INT8 layer coverage is measured from A100-built engines and used only as a graph diagnostic.

CNN–based
			FP32		FP16		INT8
Model	Par.	INT8	J/img	img/s	J/img	img/s	J/img	img/s
	(M)	%lyr †
APGCC	18.7	94	4.98	4.5	2.40	7.8	1.46	9.5
CHF	21.5	100	6.94	3.5	3.17	6.6	1.61	8.8
GCHF	21.5	100	7.01	3.5	3.18	6.6	1.59	9.0

Appendix

Table F.4 : Deployment performance on JHU-Crowd++. Energy and throughput for FP32, FP16, and INT8. Red marks INT8 regressions relative to FP16. † INT8 layer coverage is measured from A100-built engines and used only as a graph diagnostic.

Panel of Figure 39 — Figure F.1 : Runtime decomposition. Four-dataset median stage times for five models with INT8 slowdowns and four faster controls; slowdown models are shown in red , and hatched bars denote INT8. The observed regressions are concentrated in GPU execution rather than transfer or dispatch.

Panel of Figure 40 — Figure F.2 : INT8 coverage and engine diagnostics. Models are ordered by INT8 speedup over FP16; $\bullet$ denotes an INT8 slowdown. Darker shading indicates greater structural risk, with R1–R5 defined in the text.

Model	all	comp.	INT8	cov. (%)	INT8	FP16	FP32	build (s)
	layers				plan layers (%)
CHF	27	19	19	100.0	93	4	4	44
GCHF	27	19	19	100.0	93	4	4	40
CCST	349	63	63	100.0	43	47	10	136
CSRNet	23	17	17	100.0	91	4	4	11
DM-Count	34	19	19	100.0	75	21	4	14

Appendix

Table F.5 : INT8 layer census. Dataset-median A100 build-plan statistics for each model. Compute coverage is the fraction of arithmetic layers assigned to INT8; precision shares are computed over all plan layers. Values describe compiled graph structure rather than device execution.

Crowd Counting

Crowd counting estimates how many people appear in images or video, supporting occupancy monitoring, safety planning, and crowd management.

See all 23 papers

COD10K-C: Benchmarking Robustness of Camouflaged Object Detection Under Natural Image Corruptions
May 23, 2026Arafat Hossain SayemImage Segmentation Image Corruption Robustness
Generative Adversarial Perturbations with Cross-paradigm Transferability on Localized Crowd Counting
Mar 25, 2026Alabi Mehzabin Anisha, Guangjing Wang, Sriram ChellappanAdversarial Attacks Transfer-Based Adversarial Attacks
University of South Florida, Tampa, Florida, USA
RAMP: Robust Adaptive Mixed-Precision Quantization for Edge CPU Vision Models
Sep 23, 2026David Población-Criado, Dario Garcia-Gasulla, Eduardo QuinonesMixed-Precision Quantization Edge Inference
Barcelona Supercomputing Center (BSC) · Universitat Politècnica de Catalunya - BarcelonaTech (UPC)

May 23, 2026J/K move · Enter open · S save

COD10K-C: Benchmarking Robustness of Camouflaged Object Detection Under Natural Image Corruptions

Arafat Hossain Sayem

Image Segmentation Image Corruption Robustness Corruption Robustness Unsupervised Camouflaged Object Detection Camouflaged Object Detection

Open paper AskPDF arXiv

Camouflaged object detection has improved substantially, but most standard benchmarks evaluate models only on clean images. This is not realistic because real cameras often capture blur, sensor noise, weather effects, and compression artifacts. We present COD10K-C, a corruption robustness benchmark based on COD10K. It includes 8 corruption types and 5 severity levels, giving 40 conditions and 81,040 evaluation pairs in total. We evaluate three popular camouflaged object detection models, SINet-v2, PFNet, and ZoomNet, as well as a lightweight model called RobustCODLite. All models show clear performance drops on corrupted images. Motion blur and Gaussian blur cause the largest drops, with SINet-v2 losing 18.5 Dice points under motion blur. Brightness and fog are less harmful. RobustCODLite uses corruption augmentation, a frequency-prior branch, and an uncertainty-consistency loss. It retains 92.3% of its clean Dice score under corruption, compared with 87.7% for SINet-v2, 84.8% for ZoomNet, and 84.1% for PFNet. On the hardest corruptions, RobustCODLite matches or outperforms models that perform better on clean data. We will release the COD10K-C GitHub repository to support future research in robust camouflaged object detection.

Abstract

Crowd Counting

Explore similar work

COD10K-C: Benchmarking Robustness of Camouflaged Object Detection Under Natural Image Corruptions

Generative Adversarial Perturbations with Cross-paradigm Transferability on Localized Crowd Counting

RAMP: Robust Adaptive Mixed-Precision Quantization for Edge CPU Vision Models