|
OAR: Training Quantization-Friendly Object Detectors via Outlier-Aware Restriction |
Soundness: 2: fair
Presentation: 2: fair
Contribution: 2: fair
Rating: 2: reject
Confidence: 4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. |
This paper introduces Outlier-Aware Restriction (OAR), a training framework aiming to improve quantization robustness for object detectors by regularizing activations and weights. The method comprises three main components: (1) Resonant Shrinkage loss for activation stabilization, (2) Output Adaptation for branch-wise scaling, and (3) Edge-Aware KURE loss for weight regularization. Experiments on MS COCO show improvements under low-bit quantization (especially 4-bit and 2-bit) for several detectors (RetinaNet, ATSS, YOLOX-tiny).
1. The framework is conceptually simple and compatible with existing quantization methods.
2. The experime ntal section is fairly detailed, covering multiple detectors and quantization schemes.
1. Overall, the work combines existing ideas rather than presenting a clear new theoretical insight or algorithmic innovation.
2. Weak theoretical grounding. The methodology lacks a solid theoretical justification for why these specific regularizers should directly lead to better quantization robustness. The intuition about “outliers” is plausible but never formalized or empirically analyzed beyond heuristic evidence.
3. Experimental analysis lacks rigor. Improvements in mAP are relatively small and inconsistent across models (e.g., many results fluctuate within 1–2 mAP). Comparisons are limited to a few older baselines (e.g., LSQ, QDrop); more recent and stronger quantization approaches are missing.
See the weaknesses. |
Fully AI-generated |
|
OAR: Training Quantization-Friendly Object Detectors via Outlier-Aware Restriction |
Soundness: 2: fair
Presentation: 2: fair
Contribution: 2: fair
Rating: 2: reject
Confidence: 4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. |
This paper introduces OAR (Outlier-Aware Restriction), a training framework designed to improve the robustness of object detectors under low-bit quantization. It addresses activation and weight outliers by proposing Resonant Shrinkage Loss, Output Adaptation, and Edge-Aware KURE Loss, leading to more quantization-friendly models. Experiments on the MS COCO dataset demonstrate significant performance improvements, especially at 4-bit quantization, with minimal calibration data, making it effective for real-world deployment.
1. The authors highlight an interesting insight: the statistical distribution differences between the two, which could significantly impact the quantization process. This observation adds valuable depth to the understanding of model performance under quantization.
2. The method presented in the paper is clearly written and easy to understand, with well-defined formulas that effectively serve the motivation. The clarity of the approach helps convey the core ideas in a straightforward manner, making it accessible while maintaining technical rigor.
1. In Section 3.3, where the question mark appears incorrectly. It seems to be a citation mistake.
2. Wouldn't it be beneficial to include methods like SmoothQuant and QuaRot in the baseline comparison? These PTQ techniques are effective in handling outliers in large models and have shown good results in mitigating outliers during post-training quantization. Since the proposed approach is training-based and involves a longer time cost, should it not ideally achieve better or more stable performance compared to these PTQ methods? Or could OAR further enhance the performance of these methods?
3. The paper does not provide information on the training time and convergence speed of OAR. Would introducing OAR impact the model’s training convergence? Specifically, does the training time increase significantly, and does OAR affect the speed at which the model converges compared to traditional methods?
4. The experimental results show only marginal improvements in quantization performance, with some settings (e.g., QDrop with RetinaNet-ResNet50-INT2) performing better. There are also cases where the performance is worse than the FP32 baseline. What could explain these inconsistencies?
See weaknesses |
Moderately AI-edited |
|
OAR: Training Quantization-Friendly Object Detectors via Outlier-Aware Restriction |
Soundness: 2: fair
Presentation: 1: poor
Contribution: 2: fair
Rating: 2: reject
Confidence: 3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked. |
This paper proposes OAR (Outlier-Aware Restriction), a training framework to improve quantization-friendliness of object detectors. The method introduces three components: Resonant Shrinkage loss to constrain activation boundaries, Output Adaptation with learnable scaling factors for task branches, and Edge-Aware KURE loss for weight regularization. The method is evaluated on both Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) pipelines, showing improvements in various low-bit scenarios.
- Identifies an interesting problem of activation distribution discrepancies in detection heads
- Provides comprehensive ablation studies showing component contributions
- The core premise is flawed. Modifying pre-training with quantization-oriented losses is a form of model re-training. This contradicts the standard, widely-accepted definition of PTQ, which performs calibration post-training without updating model weights. The work fails to justify why this pre-training approach is preferable to directly using the proposed losses in a QAT setting.
- The paper's evaluation relies on a weakly-defined "naive PTQ" baseline. It fails to compare against established PTQ calibration methods such as OMSE, Percentile or entropy-based observers.
- The paper lacks comparisons with state-of-the-art PTQ methods (e.g., BRECQ, AdaRound) and modern detector architectures (e.g., Deformable DETR, DINO).
See Weaknesses. |
Lightly AI-edited |