Continuous Anisotropic Downscale Predictor Optimization for Bitrate-Constrained Object Detection with Bias-Calibrated End-to-End Compression Learning

초록

This paper proposes a continuous anisotropic downscale predictor (ADP) for bitrate-constrained object detection. Existing resolution-control methods often use isotropic scale factors, discrete scale candidates, or pre-generated optimal scaling factors, which limits their ability to preserve direction-dependent image structures under a strict bitrate budget. To address this limitation, the proposed ADP predicts continuous horizontal and vertical downscaling factors for each input image. This continuous prediction provides finer bitrate-accuracy adaptation than discrete resolution candidates, while differentiable anisotropic downscaling is applied before JPEG compression. The framework is trained using a detection objective and a differentiable bitrate surrogate so that resolution control is optimized with respect to both encoded bitrate and detection accuracy. To reduce the mismatch between the estimated bitrate and the actual JPEG bitstream size, we further introduce online bitrate calibration using real JPEG encoding statistics. Experiments on COCO val2017 show that the proposed method provides improved matched-bitrate detection accuracy compared with previous resolution-control methods. Across the evaluated target bitrate points, the proposed method achieves up to 8.41 percentage-point improvement in mAP@50 and 5.28 percentage-point improvement in mAP@50:95 over the compared baselines. In particular, the results on the aspect-ratio-extreme subset show that independent horizontal and vertical resolution control is beneficial for images with strong directional structure. Additional intermediate feature-similarity and efficiency analyses further support that the proposed ADP provides practical bitrate-aware resolution control for object detection. © 2013 IEEE.

키워드