ITDRCNN: An Interactive Task-Decoupled RCNN for enhanced object detection

Traditional RCNN-based object detection frameworks typically utilize a sibling head, where a classifier and a regressor share proposal features and most network parameters. However, this design often leads to a severe spatial misalignment problem, as the optimal Regions of Interest (ROIs) for different tasks, such as classification and localization, can vary significantly. To address this issue, we conducted a thorough analysis of the propagation process for a proposal, from its generation to loss calculation. The results reveal that the position, shape, and size of a proposal can be optimized during training to adapt to specific tasks. Based on this observation, we propose a novel Interactive Task-Decoupled RCNN (ITDRCNN), which disentangles the classification and localization tasks in the spatial dimension by learning two decoupled proposals. ITDRCNN contains two key mechanisms: the Decoupled Region Proposal Network (DRPN) and the Interactive Decoupled ROI Pooling (IDROIP). For each target instance, DRPN outputs two task-specific proposals. IDROIP pools RoI features from each proposal, and the resulting features are sent to separate heads for classification and localization. More importantly, IDROIP allows backpropagation for the coordinates of the proposals, ensuring that the decoupled proposals can be directly learned to adapt to different tasks during training. Thus, the spatial misalignment problem is effectively addressed. The experimental results on the MSCOCO and PASCAL VOC benchmarks demonstrate that ITDRCNN is both effective and efficient.