Deep Equilibrium Object Detection and Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Query-based methods utilize iterative refinement decoders for object detection and segmentation.
• These methods represent object instances using learnable queries that are progressively refined.

Purpose of the Study:

• Introduce a novel query-based object decoder design using deep equilibrium models.
• Enhance query refinement through implicit layer fixed-point solving.

Main Methods:

• Model query vector refinement as a fixed-point solution of an implicit layer using a two-step unrolled equilibrium equation.
• Incorporate refinement awareness during training with inexact gradient back-propagation (RAG).
• Employ a deep supervision scheme with refinement-aware perturbation (RAP) for training stability and generalization.

Main Results:

• DEQDet, an object detector based on DEQ-Decoder, shows faster convergence, reduced memory consumption, and superior performance compared to AdaMixer.
• Achieved 49.6 mAP and 33.9 APs on MS COCO with ResNet50 and 300 queries (2x training).
• DEQSeg demonstrates improved box mAP and competitive mask metrics in instance segmentation.

Conclusions:

• The DEQ-Decoder offers an effective approach for query-based object detection and instance segmentation.
• The proposed methods (DEQDet and DEQSeg) achieve state-of-the-art results with improved efficiency.
• Publicly available code and models facilitate further research and application.