Answers are grounded in this paper and cite their sources.
In the design of object detection algorithms, the label Assignment and positive/negative sample identification are very important, affecting the convergence speed and object detection performance. Nowadays, positive/negative sample identification mainly focuses on one-to-many label assignment of Ground-Truth and Bounding Box, and few researchers have paid attention to the label assignment of end-to-end object detection models. In this paper, aiming at end-to-end object detection models, we propose a dynamic label assignment algorithm based on the Hungarian matching method, and normalization methods of L1 loss at the stage of cost calculation and loss calculation. We test the proposed dynamic label assignment algorithm included the end-to-end model SparseRCNN with the COCO dataset. The resulting AP achieved an improvement of 0.4 when we train the model on four GPUs and an improvement of 0.6 when training the model on a single GPU.
No takes yet. Share an insight, caveat, or question.
A 2022 study studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: