acceptodds
Withdrawn from ICLR 2027

AxisQ: Depth-Adaptive Channel–Frequency Quantization for Camouflaged Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Camouflaged object detection (COD) depends on weak texture and boundary cues that can be destroyed by aggressive post-training quantization (PTQ). Existing PTQ methods mainly optimize quantizers or reconstruct network blocks, while using a single precision-allocation domain throughout hierarchical backbones. Instead, we introduce AxisQ, a depth-adaptive bit-allocation framework motivated by the observation that the useful allocation axis changes with feature depth. Under a fixed per-layer average activation-bit budget, AxisQ allocates precision in the channel domain for shallow features and in the local-frequency domain for deep features, using COD-aware sensitivity to better preserve subtle foreground regions and ambiguous boundaries. We also formulate a boundary-aware Hessian approximation for reconstruction by incorporating boundary-weighted gradient and error statistics. Experiments across multiple COD models and diverse benchmarks demonstrate improved accuracy over the strongest baselines at low bit widths, with relative weighted F-measure gains of 12.7% for CFRN and 33.3% for CamoFormer-P on CAMO under W2A2 quantization. These results support adapting the allocation domain to feature depth as a transferable quantization strategy for hierarchical COD models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Post-training Quantization, Camouflaged Object Detection, Vision Transformers

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fVsBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AxisQ, a post-training quantization framework for COD. Its core claim is that the useful allocation axis for bit precision changes with feature depth. Shallow stages benefit more from channel-domain allocation, while deep stages from local-frequency domain allocation. AxisQ realizes this via task-sensitive channel costs (empirical-Fisher), COD-aware frequency costs (COD-GGN), and an auxiliary boundary-aware Hessian reconstruction. Experiments span 2 backbones across 4 benchmarks with 5 precision settings and compare with multiple PTQ baselines. The results shows the largest gains under W2A2 setting, reaching 12.7% weighted-F improvement on CFRN.

Critical strengths
  1. The improvements are uniform and consistent across all settings and achieve large margins on the hardest setting.
  2. Treating per-channel and per-frequency not as a fixed setting but as a depth-adaptive choice is a modest but real contribution.
Critical weaknesses
  1. The central motivation is weakly supported. The sensitivity concentrates in channels is evidenced only by Figure 2. The figure only convers one selected COD image without any dataset-level aggregation. Therefore the generalizable claim is stated far beyond its evidence.
  2. The complex components seem contribute near-zero gain. COD-GGN occupies a lot 篇幅, yet accounts for only 0.008 in Table 3. The boundary-aware Hessian is worse. The gains are carried by the conceptually simple mixed-domain, while the two complex components add little measurable benefit despite their heavy derivation.
Questions for authors
  1. Can you provide some statistics for the channel/frequency concentration in Figure 2? For example the distribution, mean and variance across COD10k dataset?
  2. The largest gains are at W2A2, but deployment timing covers only W4A4. Is there any hardware path on which 2-bit activations yield real speedup, or is W2A2 purely a simulation result?
  3. Given Table 6 shows no FβmF_\beta^m change, what is the intended justification for retaining the boundary-aware Hessian component?
  4. The method uses many fixed hyperparameters (γ=0.30, β=0.10, α=2, η=0.1, brefb_ref=2). How sensitive are results to them?
Reviewer_ga9pOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AxisQ, a depth-adaptive post-training quantization framework for camouflaged object detection. The main idea is to allocate activation precision across channels in shallow network stages and across local-frequency bands in deeper stages. The method uses empirical-Fisher sensitivity for channel-wise allocation and combines reconstruction error with a COD-aware generalized Gauss–Newton approximation for frequency-wise allocation. The authors also introduce an auxiliary boundary-aware Hessian approximation for block reconstruction. Experiments are conducted on CFRN and CamoFormer-P across four COD benchmarks under different quantization settings. The results show consistent improvements over several existing PTQ methods, particularly under W2A2 quantization. Additional ablation studies examine the contributions of different allocation strategies and task-aware sensitivity estimation.

Critical strengths
  1. The paper introduces an interesting approach to precision allocation by considering the different characteristics of shallow and deep features. Instead of applying the same allocation domain throughout the network, AxisQ combines channel-wise and frequency-wise allocation within a unified framework. This is a reasonable design for hierarchical models and provides a useful direction for improving low-bit quantization.
  2. The experimental results are promising, particularly under aggressive W2A2 quantization. AxisQ consistently outperforms the compared methods on both CFRN and CamoFormer-P across four COD datasets. The improvements are supported by multiple evaluation metrics, and the ablation results further indicate that combining the two allocation domains is more effective than using either domain alone.
Critical weaknesses
  1. The paper argues that shallow features benefit more from channel-wise allocation, while deeper features are better suited to frequency-wise allocation. However, Figure 2 presents sensitivity analysis on a selected example, leaving the generality of this observation unclear. Although Table 3 shows that the mixed-domain strategy outperforms single-domain alternatives, it does not sufficiently justify the specific stage division adopted in AxisQ. As a result, the central motivation for depth-adaptive allocation remains insufficiently supported.
  2. AxisQ maintains a fixed average activation-bit budget, but this does not necessarily imply equivalent computational or memory costs. The proposed frequency-domain allocation also requires DCT/IDCT operations and heterogeneous bit-width processing, which may introduce additional overhead. While Appendix E provides deployment results under W4A4, the practical efficiency under W2A2, where the largest performance gains are reported, remains unclear. This limits the evidence supporting the method's practical deployment advantages.
Questions for authors
  1. How sensitive is AxisQ to the choice of stage division? How does the proposed allocation strategy compare with alternative stage divisions or reversed channel-frequency assignments?
  2. Are the depth-dependent sensitivity patterns shown in Figure 2 consistent across different samples and both backbones? What evidence supports the generality of these observations?
  3. What are the actual model size, inference latency, and memory requirements of AxisQ under W2A2 quantization compared with the baselines? How much computational overhead is introduced by DCT/IDCT and mixed-precision operations?
Reviewer_uUJWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AxisQ, a post-training quantization method for camouflaged object detection (COD). It allocates activation bits across channels in shallow stages and DCT frequency bands in deeper stages, guided by COD-aware sensitivity. It also introduces a boundary-weighted Hessian for reconstruction.

Critical strengths

Consistent improvements at very low bit widths. At W2A2, AxisQ outperforms strong reconstruction-based baselines across two backbones and four benchmarks. For example, the weighted F-measure on CAMO improves from 0.645 to 0.727 for CFRN. The COD-GGN cost is also efficiently derived, requiring only one VJP probe per layer. The paper additionally includes a deployment experiment, which is relatively uncommon in this area.

Critical weaknesses
  1. The core motivation lacks sufficient evidence. The claim that the optimal bit-allocation domain varies with network depth is supported mainly by a single selected image (Fig. 2). Moreover, the channel/frequency split is manually fixed (channels for stages 1–2 and frequency bands for stages 3–4), rather than adaptively determined. An ablation on different split points or an automatic per-layer selection strategy would better support the motivation.

  2. The baseline comparisons are not fully fair. AxisQ allows each channel or frequency band to use 0–6 bits under an average bit budget, introduces DCT transforms, and uses labeled calibration data, while the baselines use uniform precision. Comparisons with existing mixed-precision allocators under the same budget, as well as a DCT-based variant with uniform bit widths, are needed to isolate the proposed contributions.

  3. Practical deployment remains unclear. Per-channel and per-band mixed precision, together with DCT/IDCT operations at every layer, may be difficult to implement efficiently with integer kernels. The W4A4 integer implementation achieves only 1.07× speedup over FP32, while the 1.52× variant uses FP16. W2A2, where the largest gains are reported, is not evaluated on hardware. Memory savings are also missing.

  4. The remaining contributions show limited gains. Boundary-aware reconstruction is evaluated only on CHAMELEON at W2A4 under a separate protocol, improving weighted F-measure by just 0.003. At W2A4 and W4A4, most per-dataset improvements are within 0.005, with no variance reported across calibration seeds.

Questions for authors

1.How does AxisQ compare with existing mixed-precision allocators under the same average bit budget? 2. How sensitive is performance to the channel/frequency stage split? Could the allocation domain be selected automatically for each layer? 3. Can the authors report actual latency and memory usage at W2A2, and explain how per-band mixed precision could be supported by integer kernels? 4. Does the method generalize to other dense prediction tasks, such as salient object detection or semantic segmentation?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.