TA-BiDet: Task-aligned binary object detector

Binary CNN-based object detector, largely saving storage and computation costs, has received high attention recently for the potention of efficient deployment on resource-constrained devices. However, previous designs often suffer from significant accuracy loss when compared to their real-value counterparts. This study demonstrates that the primary reason lies in the mis-alignment of the classification and regression tasks, which is generally caused by the limited representational capacity of the binary neural network and biased training procedures used in traditional detection frameworks. Based on this observation, we propose TA-BiDet (Task-Alignment Binary object Detection) that can guarantee aligned training of the two tasks by adopting a task-aware feature disentanglement (TFD) network architecture with an alignment-oriented learning (AOL) approach. The proposed approaches can ensure more informative and tailored task-specific features to be learned jointly for each task and select the most accurate detected boxes with both high confidence scores and precise locations. Experiments on the PASCAL VOC and COCO datasets have shown that TA-BiDet outperformed state-of-the-art binary object detectors by a considerable margin. Moreover, TA-BiDet has successfully narrowed the performance gap with the real-valued SSD300 detector to only 0.7% in terms of mAP, and reduced the model size by 4.86× and the total OPs by 9.89×, respectively.