Summary of Deim: Detr with Improved Matching For Fast Convergence, by Shihua Huang et al.
DEIM: DETR with Improved Matching for Fast Convergence
by Shihua Huang, Zhichao Lu, Xiaodong Cun, Yongjun Yu, Xiao Zhou, Xi Shen
First submitted to arxiv on: 5 Dec 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A novel training framework called DEIM accelerates convergence in real-time object detection with Transformer-based architectures, specifically DETR models. To address the limitations of one-to-one matching in these models, DEIM employs a Dense O2O matching strategy that increases positive samples per image using standard data augmentation techniques. This approach not only speeds up convergence but also introduces low-quality matches that could impact performance. To optimize these matches, the Matchability-Aware Loss (MAL) is proposed as a novel loss function that enhances the effectiveness of Dense O2O. The efficacy of DEIM is validated through extensive experiments on the COCO dataset, which consistently boost performance while reducing training time by 50%. When paired with RT-DETRv2, DEIM achieves 53.2% AP in just one day of training on an NVIDIA 4090 GPU. Additionally, DEIM-trained real-time models outperform leading detectors, achieving 54.7% and 56.5% AP at 124 and 78 FPS respectively, without requiring additional data. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary DEIM is a new way to train Transformer-based models for object detection that makes them better and faster. It solves a problem called one-to-one matching by using more information from the images. This makes the model learn faster and get better results. DEIM also has a special loss function that helps it make better decisions. The results show that DEIM is very good at detecting objects in real-time, even when compared to other state-of-the-art models. |
Keywords
» Artificial intelligence » Data augmentation » Loss function » Object detection » Transformer
