Research map / Architectures and efficiency / Attention and transformers
Deeply Tensor Compressed Transformers for End-to-End Object Detection
Peining Zhen, Ziyang Gao, Tianshu Hou, Yuan Cheng, Hai-Bao Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Tensor-train compress FFN weights and add GMHA | DETR with ResNet-50 · COCO val | AP: 40.1 -> 39.5 | Hurt |
| Add GMHA to vanilla DETR (ablation) | DETR · COCO val | AP: 40.1 -> 41.7 | Improved |
| Add GMHA to tensorized DETR (ablation) | T-DETR · COCO val | AP: 37.6 -> 39.5 | Improved |
| Apply GMHA only to encoder self-attention rather than all attention modules (ablation) | T-DETR · COCO val | AP: 39.2 -> 39.5 | Improved |
| Increase tensor-train rank from 3 to 4 (ablation) | Tensorized DETR FFN · COCO val | AP: 36.9 -> 37.6 | Improved |
| Increase tensor-train rank from 4 to 5 (ablation) | Tensorized DETR FFN · COCO val | AP: 37.6 -> 37.7 | Improved |
| Replace tensor-train decomposition with Tucker decomposition (ablation) | DETR FFN · COCO val | AP: Tucker rank 2: 37.4; rank 3: 37.7; Tucker-2 rank 4: 37.6 | Mixed |
| Quantize the ResNet-50 backbone to 8-bit | T-DETR · COCO val | AP: 39.5 -> 39.5 | No clear effect |
| Quantize the ResNet-50 backbone to 4-bit | T-DETR · COCO val | AP: 39.5 -> 37.9 | Hurt |
| Quantize ResNet-50 weights to 8-bit | ResNet-50 · ImageNet-1k val | top-1 accuracy: 76.15 -> 76.18 | No clear effect |
| Quantize ResNet-50 weights to 4-bit | ResNet-50 · ImageNet-1k val | top-1 accuracy: 76.15 -> 75.51 | Hurt |
| Use tensor-train rank 4 rather than an uncompressed FFN (ablation) | DETR FFN · COCO val | FFN parameter count: 1,048,576 -> 2,176 | Improved |
Closest papers by meaning
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.89)
- CF-DETR: Coarse-to-Fine Transformers for End-to-End Object Detection (AAAI 2022 · similarity 0.88)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.86)
- KD-DETR: Knowledge Distillation for Detection Transformer with Consistent Distillation Points Sampling (CVPR 2024 · similarity 0.86)
- AQ-DETR: Low-Bit Quantized Detection Transformer with Auxiliary Queries (AAAI 2024 · similarity 0.86)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
