Foreground-Specialized Model Imitation for Instance Segmentation

Dawei Li, Wenbo Li, Hongxia Jin; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 618-633


Instance segmentation is formulated as a multi-task learning problem. However, knowledge distillation is not well-suited to all sub-tasks except the multi-class object classification. Based on such a competence, we introduce a lightweight foreground-specialized (FS) teacher model, which is trained with foreground-only images and highly optimized for object classification. Yet, this leads to discrepancy between inputs to the teacher and student models. Thus, we introduce a novel Foreground-Specialized model Imitation (FSI) method with two complementary components. First, a reciprocal anchor box selection method is introduced to distill from the most informative output of the FS teacher. Second, we embed the foreground-awareness into student's feature learning via either adding a co-learned foreground segmentation branch or applying a soft feature mask. We conducted an extensive evaluation against the others on COCO and Pascal VOC.

Related Material

@InProceedings{Li_2022_ACCV, author = {Li, Dawei and Li, Wenbo and Jin, Hongxia}, title = {Foreground-Specialized Model Imitation for Instance Segmentation}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {618-633} }