Inlier-Centric Post-Training Quantization for Object Detection Models

TL;DR

This paper tackles the difficulty of quantizing object detection models under low-bit constraints by isolating task-relevant inliers from anomalies in activations. It introduces InlierQ, which uses gradient-based volume saliency to model an inlier–anomaly posterior via a two-component EM mixture and optimizes quantization over the inlier set, mitigating anomaly-induced distribution skew. The method yields consistent gains on COCO and nuScenes for both 2D and 3D detectors, with pronounced improvements under 4-bit activations, demonstrating improved robustness and efficiency for on-device deployment. The key contributions include a principled inlier–anomaly decomposition, a gradient-informed saliency model, and EM-based posterior inference that together enable drop-in PTQ with minimal calibration data and strong empirical performance gains.

Abstract

Object detection is pivotal in computer vision, yet its immense computational demands make deployment slow and power-hungry, motivating quantization. However, task-irrelevant morphologies such as background clutter and sensor noise induce redundant activations (or anomalies). These anomalies expand activation ranges and skew activation distributions toward task-irrelevant responses, complicating bit allocation and weakening the preservation of informative features. Without a clear criterion to distinguish anomalies, suppressing them can inadvertently discard useful information. To address this, we present InlierQ, an inlier-centric post-training quantization approach that separates anomalies from informative inliers. InlierQ computes gradient-aware volume saliency scores, classifies each volume as an inlier or anomaly, and fits a posterior distribution over these scores using the Expectation-Maximization (EM) algorithm. This design suppresses anomalies while preserving informative features. InlierQ is label-free, drop-in, and requires only 64 calibration samples. Experiments on the COCO and nuScenes benchmarks show consistent reductions in quantization error for camera-based (2D and 3D) and LiDAR-based (3D) object detection.

Figures (5)

• Figure 1: Effects of Inlier-Centric Quantization. (Top) InlierQ computes a gradient-aware volume saliency score, estimates an inlier probability map, and samples high-probability volumes to form a task-relevant inlier set (green points) around target objects. (Bottom) Kernel density estimates with scattered activations and the relative skew show that anomalies broaden and skew the activation distribution. By restricting quantization to inliers, InlierQ yields a compressed activation range while preserving task-relevant information. This leads to lower layer-wise MSE, slower relative error accumulation, and up to 3.2% mAP improvement over a baseline method li2021brecq.
• Figure 2: 3D points (left), predicted heatmap (middle), and heatmap top-$K$ overlaid on 3D points (right). In constructing the loss function \ref{['eq:heatmap_topk']}, the top-$K$ heatmap predictions are exploited to prioritize preserving model performance in object searching.
• Figure 3: Cross-modal Behavior of Gradients and Volume Saliency Scores. (Left) Gradient-domain histograms indicate that each modality has different inlier distributions. (Middle) Yet, in the space of volume saliency, both modalities exhibit a consistent, modality-invariant distribution. (Right) Volume saliency scores overlaid on the labeled input signals highlight objects while appropriately modulating the contribution of background, serving as a task-relevant scoring function.
• Figure 4: For varying threshold strictness. X-axis corresponds to the threshold parameter in \ref{['eq:inlier_set']}, and Y-axis is task performance. Inlier (or Anomaly) set defined with a higher $\tau$ applies a stricter (or less strict) decision, causing performance gain (or drop).
• Figure 5: For varying Heatmap Top-K. X-axis indicates the number of heatmap Top-K parameter, and Y-axis is the task performance.