Efficient Object Detection in Compressed Domain by Exploiting Knowledge Distillation from Pixel Domain

The proliferation of high-definition video data necessitates highly efficient processing pipelines for real-time edge analytics. However, traditional object detection architectures rely exclusively on pixel-domain inputs, which renders the computationally prohibitive decoding phase a latency bottleneck. In this paper, we propose a novel dual-phase framework designed to achieve fast and efficient object detection directly within the partially decoded compressed-domain data. First, we introduce a partial decoding paradigm featuring the Low-Frequency Spectral Prioritization method on the encoder side. By systematically discarding high-frequency residual coefficients and retaining only a sparse subset of fundamental spatial frequencies, this method dramatically reduces transmission payloads and accelerates the standard decoding process. Second, to recover the structural fidelity lost due to the intentional omission of residual data, we employ a multi-granularity cross-domain knowledge distillation architecture. This strategy aligns global contextual features, foreground boundary attention maps, and final response logits, transferring rich representational capacities from a high-performing pixel-domain teacher network to a lightweight compressed-domain student network. Comprehensive experiments utilizing RetinaNet, FCOS, and GFL object detection networks on the COCO-mini dataset demonstrate the superiority of the proposed framework. By retaining fundamental residual coefficients within the HEVC pipeline, the proposed method reduces average decoding latency while improving the mAP score by +0.96% over the conventional fully decoded pixel-domain baseline on the COCO-mini dataset.