LMVR: Lightweight multitask visual recognition with Cross Residual Log-likelihood Estimation
摘要
Currently, Multi-task deep learning algorithms have been applied in object detection, instance segmentation, and human pose estimation, etc. However, these models rely mainly on robust backbone and large-scale feature inputs to improve model performance, but lack a unified learning approach. This paper aims to improve the original model’s accuracy by leveraging lightweight backbone and a unified multitasking learning approach, while avoiding the introduction of additional parameters. Therefore, this paper proposes an auxiliary learning called Lightweight Multitasking Visual Recognition (LMVR). This method utilizes prior knowledge to determine the distribution of multitask learning, that is, towards a Gaussian distribution. First, a Gaussian heat map module called Keypoint Processing Estimation (KPE) is incorporated into the backbone, so as to supervise the intermediate learning process during training; second, a polar coordinate transformation is introduced to replace the original 4-vector system to reduce parameters; Finally, a Cross Residual Log-likelihood Estimation (CRLE) loss is proposed to address the issue of exponentially increasing parameters when using RLE loss for large-scale image feature calculation. Experimental results demonstrate that the LMVR network achieves a box Average Precision (AP) of 46.2 (Our3) and a keypoint AP of 46.6 (Our4). Compared to LSNet, Our1 improves by 6.2