Post-training W4A4 for camouflaged object detection

COD–TDQ

When W4A4 Breaks Camouflaged Object Detection

Token-Group Dual-Constraint Activation Quantization

4-bit weights are not the breaking point. 4-bit activations are. COD-TDQ isolates the token-local failure and preserves the weak boundary evidence that camouflage detectors need.

Tianqi Li¹ Wenyu Fang¹ Xin He² Xue Geng³ Xu Cheng² Yun Liu^1,4,5,*

¹ Nankai University · ² Tianjin University of Technology · ³ A*STAR · ⁴ AAIS, Nankai University · ⁵ Nankai International Advanced Research Institute, Shenzhen Futian

PaperPDF ↗ arXivv2 ↗ CodeGitHub ↗ DatasetHF ↗ WeightsHF ↗

THE W4A4 CLIFFCFRN · NC4K · S_α

FP320.888

Naive W4A40.443

COD-TDQ0.837

The mask does not merely soften—it nearly disappears. COD-TDQ restores the signal without retraining.

41.6 → 14.2%zeroed boundary activations

+11.8–16.1 ptsS_α over strongest CFRN W4A4 baselines

1.50×measured throughput vs. FP32

108.19 MBCFRN deployment artifact

01 / THE FAILURE

W4A4 is not a gentle degradation.

Holding 8-bit activations nearly preserves FP32 accuracy. Moving activations to 4 bits causes a task-specific cliff.

CFRN · NC4KPaper Table 1

FP32baseline

.888S_α

W8A8naive PTQ

.887S_α

W4A84-bit weights

.882S_α

W4A44-bit activations

.443S_α

Weights can go to 4 bits while the mask survives. The cliff appears when activations do.

Across all four CFRN datasets, naive W4A8 averages 0.8776 S_α, close to FP32 at 0.8864. Naive W4A4 falls to 0.4448. This isolates activation precision as the dominant failure factor.

WHY DOES W4A4 BREAK COD?

A shared range silences the signal COD needs most.

Follow the failure from a background spike to a missing boundary.

Heavy-tailed background

A few background tokens carry extreme activation spikes.

Shared-range domination

The outlier dictates one clipping radius c for every token.

Coarse 4-bit step

The wider range inflates Δ = c / 7 for every token.

Boundary falls to zero

Responses inside |x| ≤ Δ/2 are rounded to exactly zero.

Paper Figure 1: radar diagnostics and qualitative mask comparison for FP32, naive W4A4, and COD-TDQ — **Figure 1.** On CFRN/NC4K, naive W4A4 widens the normalized clipping range, inflates Δ, and raises zero-bin mass ρ₀; the target mask collapses with it.

02 / COD-TDQ

Localize the range. Then constrain it.

COD-TDQ is a purely post-training, token-local activation quantizer. Its two components solve different parts of the same collapse loop—and only their combination consistently restores COD masks.

SHARED RANGE

one background token sets Δ for all

01DSTGDirect-Sum Token-Group

02DCRPDual-Constraint Range Projection

W4A4

weak structure remains signed

REMOVE CROSS-TOKEN COUPLING

Direct-Sum Token-Group

DSTG partitions each token vector into fixed-size channel groups and gives every group its own activation range. A heavy-tailed background group can no longer dictate the resolution used by unrelated boundary groups.

x = ⊕_k x_kgroup size g = 32 in all reported experiments

KEEP EACH GROUP STABLE

Dual-Constraint Range Projection

DSTG removes interference between groups, but a single group can still be heavy-tailed. DCRP projects its clipping radius against two explicit stability bounds.

η_k = Δ_k / σ_k ≤ τBound step size relative to group dispersion.

ρ_0,k ≤ zrBound the fraction that collapses into zero.

Paper method overview: token extraction from Transformer layers, DSTG grouping, DCRP constraints, and INT4 quantization — **Method overview.** COD-TDQ applies token-group range selection before uniform signed W4A4 quantization. The projection uses one shared setting across datasets and both backbones: g = 32, τ = 1.0, zr = 0.2.

03 / MECHANISM EVIDENCE

The diagnostics move in the predicted direction.

Two paper figures test the mechanism directly: DSTG reduces scale interference and boundary zeroing; DCRP reduces violations of both stability bounds.

Token-wise range visualization and boundary activation scatter before and after COD-TDQ — Token-group scaling suppresses background-driven range inflation. The fraction of zeroed boundary activations falls from **41.6%** under naive W4A4 to **14.2%** with COD-TDQ.

Distributions and constraint regions for eta and zero-bin mass before and after DCRP — Groups exceeding C1 drop from **72.60% to 0.00%**. Groups with pre-projection ρ₀ above zr drop from **98.36% to 20.87%** under COD-TDQ statistics.

04 / QUANTITATIVE RESULTS

Recovery holds across models and datasets.

All rows below use the paper's common evaluation pipeline. COD-TDQ and PTQ baselines are W4A4; FP32 is shown as a reference. S_α is higher-better, MAE is lower-better.

CFRN results from main paper Table 2
Method	CAMO		CHAMELEON		COD10K		NC4K
Method	S_α	MAE	S_α	MAE	S_α	MAE	S_α	MAE
FP32 reference	.876	.042	.912	.019	.870	.022	.888	.030
Naive W4A4	.418	.182	.447	.141	.471	.093	.443	.151
RepQ-ViT strongest S_α baseline	.676	.099	.701	.067	.676	.062	.718	.072
COD-TDQ	.813	.070	.862	.040	.802	.038	.837	.052

COD-TDQ improves S_α by 11.8–16.1 points over the strongest CFRN W4A4 baseline on each dataset. Full five-metric results and all baselines are in the paper.

ESCNet results from supplementary Table S2
Method	CAMO		CHAMELEON		COD10K		NC4K
Method	S_α	MAE	S_α	MAE	S_α	MAE	S_α	MAE
FP32 reference	.8755	.0408	.8985	.0226	.8734	.0204	.8929	.0278
Naive W4A4	.4680	.1687	.4505	.1400	.5455	.0808	.5762	.1201
RepQ-ViT strong baseline	.7707	.0786	.8006	.0467	.7755	.0422	.8182	.0515
COD-TDQ	.8528	.0502	.8899	.0234	.8539	.0245	.8813	.0312

COD-TDQ restores ESCNet to near-FP32 W4A4 accuracy on all four benchmarks; numbers are reported at full precision in supplementary Table S2.

COMPONENT CHECK · NC4K

Neither component tells the whole story alone.

Naive W4A4.443

DSTG only.451

DCRP only.435

DSTG + DCRP.837

On CFRN, local ranges alone remain vulnerable to heavy tails inside each group; projection alone cannot remove cross-token coupling everywhere. Together they recover the mask.

05 / QUALITATIVE RESULTS

What the numbers mean: the object comes back.

Weak contours, textured backgrounds, and low-contrast regions expose the failure most clearly. COD-TDQ tracks FP32 structure where naive W4A4 becomes fragmented or nearly blank.

Three challenging camouflaged scenes with ground truth and predictions from COD-TDQ, FP32, naive W4A4, RepQ-ViT, IGQ-ViT and PTQ4SAM — Each example uses two rows for CFRN and ESCNet. COD-TDQ retains thin structure and complete foreground regions; naive W4A4 frequently produces sparse fragments.

06 / MEASURED DEPLOYMENT

Accuracy recovery does not erase the low-bit gain.

The paper benchmarks end-to-end CFRN inference on one NVIDIA RTX 4090 at 384 × 384, batch size 4, using the released Triton-based quantized path.

1× RTX 4090384² inputbatch 4CFRN

THROUGHPUT44.21images / second

1.50× vs. FP32

LATENCY22.61ms / image

ARTIFACT108.19MB

FP32: 775.40 MB

PEAK GPU MEMORY834.92MB

FP32: 1421.31 MB

CFRN runtime efficiency from main paper Table 5
Method	Weight / artifact ↓	FPS ↑	Latency ↓	Peak memory ↓
FP32	775.40 MB	29.40	34.02 ms/img	1421.31 MB
Naive W4A4	108.30 MB	43.95	22.75 ms/img	843.30 MB
RepQ-ViT	108.15 MB	44.18	22.64 ms/img	844.67 MB
COD-TDQ	108.19 MB	44.21	22.61 ms/img	834.92 MB

07 / ABSTRACT

The paper, in brief.

Camouflaged object detection depends on evidence that aggressive activation quantization is unusually likely to erase.

COD-TDQ diagnoses a task-specific W4A4 cliff: heavy-tailed background tokens dominate a shared activation range, enlarge the quantization step, and push weak but structured boundary cues into the zero bin. Direct-Sum Token-Group scaling removes cross-token range domination; Dual-Constraint Range Projection bounds the step-to-dispersion ratio and zero-bin mass within each group. Across CAMO, CHAMELEON, COD10K, and NC4K with CFRN and ESCNet, the method improves W4A4 accuracy without retraining.

08 / CITATION & RESOURCES

Build on COD-TDQ.

Paper, code, datasets, released FP32 and quantized checkpoints, plus Chinese documentation are publicly linked below.

Paper / arXiv v2↗ Code / GitHub↗ COD datasets / Hugging Face↗ Model weights / Hugging Face↗ Chinese documentation→ Chinese paper PDF→

@inproceedings{li2026codtdq,
  title={When W4A4 Breaks Camouflaged Object Detection:
         Token-Group Dual-Constraint Activation Quantization},
  author={Li, Tianqi and Fang, Wenyu and He, Xin and
          Geng, Xue and Cheng, Xu and Liu, Yun},
  booktitle={Proceedings of the European Conference
             on Computer Vision},
  year={2026}
}