Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
[CVPR'2022] SAM-DETR & SAM-DETR++: Official PyTorch Implementation
| Date | Stars |
|---|---|
| 2026-07-24 | 300 |
| 2026-07-25 | 300 |
| 2026-07-28 | 300 |
| 2026-07-30 | 300 |
| 2026-08-06 | 300 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# SAM-DETR (Semantic-Aligned-Matching DETR) [](https://arxiv.org/abs/2203.06883) [](https://github.com/dk-liang/Awesome-Visual-Transformer) [](https://GitHub.com/Naereen/StrapDown.js/graphs/commit-activity) [](http://makeapullrequest.com) [](https://github.com/ZhangGongjie/SAM-DETR/blob/master/LICENSE) This repository is an official PyTorch implementation of the CVPR 2022 paper "[Accelerating DETR Convergence via Semantic-Aligned Matching](https://arxiv.org/abs/2203.06883)". <b>*[UPDATE on 21 Apr 2022]*</b> We found that with a very simple modification (with no extra computational cost), SAM-DETR can achieve better performance. On MS-COCO, **SAM-DETR w/ SMCA** can achieve **37.0 AP** within 12 epochs, and **42.7 AP** within 50 epochs. We will release the updated training scripts, model weights, and logs in the future. Please stay tuned! ## Introduction <b> TL;DR </b> SAM-DETR is an efficeint DETR-like object detector that can converge wihtin 12 epochs and outperform the strong Faster R-CNN (w/ FPN) baseline. The recently developed DEtection TRansformer (DETR) has established a new object detection paradigm by eliminating a series of hand-crafted components. However, DETR suffers from extremely slow convergence, which increases the training cost significantly. We observe that the slow convergence can be largely attributed to the complication in matching object queries to encoded image features in DETR's decoder cross-attention modules. <div align=center> <img src='.assets/matching_complication.jpg' width="70%"> </div> Motivated by this observation, in our paper, we propose SAM-DETR, a Semantic-Aligned-Matching DETR that can greatly accelerates DETR's convergence without sacrificing its accuracy. SAM-DETR addresses the slow convergence issue from two perspectives. First, it projects object queries into the same embedding space as encoded image features, where the matching can be accomplished efficiently with aligned semantics. Second, it explicitly searches salient points with the most discriminative features for semantic-aligned matching, which further speeds up the convergence and boosts detection accuracy as well. Being like a plug and play, SAM-DETR complements existing convergence solutions well yet only introduces slight computational overhead. Experiments show that the proposed SAM-DETR achieves superior convergence as well as competitive detection accuracy. At the core of SAM-DETR is a plug-and-play module named "Semantics Aligner" appended ahead of the cross-attention module in DETR's each decoder layer. It also models a learnable reference box for each object query, whose center location is used to generate corresponding position embeddings. <div align=center> <img src='.assets/decoder_layer.jpg' width="90%"> </div> The figure below illustrates the architecture of the appended "Semantics Aligner", which aligns the semantics of "encoded image features" and "object queries" by re-sampling features from multiple salient points as new object queries. <div align=center> <img src='.assets/semantics_aligner.jpg' width="78%"> </div> Being like a plug-and-play, our approach can be easily integrated with existing convergence solutions (*e.g.*, SMCA) in a complementary manner, boosting detection accuracy and convergence speed further. Please check [our CVPR 2022 paper](https://arxiv.org/abs/2203.06883) for more details. ## Installation ### Pre-Requisites You must have NVIDIA GPUs to run the codes. The implementation codes are develo
Excerpt of 16,105 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:e61daf8b39a403e8, topic:computer-vision, topic:object-detection, readme:object detection
matched fp:e61daf8b39a403e8, topic:deep-learning, topic:pytorch
matched fp:e61daf8b39a403e8, topic:transformer, readme:model weights