What question did this study set out to answer?

To accelerate the training convergence and enhance the accuracy of real-time Transformer-based object detection architectures by addressing sparse supervision.

June 10, 2025

DEIM: DETR with Improved Matching for Fast Convergence

View Full Paper

Ask AI

Bookmark

Authors

SHShihua HuangInnovation Cluster (Canada)ZLZhichao LuCity University of Hong Kong XCXiaodong CunGreat Bay University

Discussion

Member takes

Implication

Benchmarking study demonstrates accelerated training and superior accuracy in real-time object detectors, indicating enhanced efficiency for vision transformers.

Key Points

To accelerate the training convergence and enhance the accuracy of real-time Transformer-based object detection architectures by addressing sparse supervision.
Introduced a Dense one-to-one (O2O) matching strategy that increases positive supervision per image using augmented targets.
Formulated Matchability-Aware Loss (MAL) to suppress low-quality matches introduced by dense target generation.
Benchmarked on the COCO dataset by integrating the framework into RT-DETR, RT-DETRv2, and D-FINE architectures using NVIDIA RTX 4090 and T4 GPUs.
Reduced total detector training time by 50% across evaluated architectures.
Achieved 53.2% average precision (AP) on RT-DETRv2 within a single day of training on one NVIDIA RTX 4090 GPU.
Yielded 54.7% AP at 124 FPS (DEIM-D-FINE-L) and 56.5% AP at 78 FPS (DEIM-D-FINE-X) on an NVIDIA T4 GPU without extra training data.

Cite This Study

Huang et al. (2025) studied this question.

synapsesocial.com/papers/69d835db05ee2ba81dbef379 https://doi.org/10.1109/cvpr52734.2025.01412

View Full Paper

Ask AI

Bookmark

Implication

Benchmarking study demonstrates accelerated training and superior accuracy in real-time object detectors, indicating enhanced efficiency for vision transformers.

DEIM: DETR with Improved Matching for Fast Convergence

Discussion

Member takes

Implication

Key Points

Cite This Study

DEIM: DETR with Improved Matching for Fast Convergence

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider