Research map / Detection and segmentation / CNN features
An Empirical Study of Adder Neural Networks for Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| unfreeze BN statistics in pre-trained adder backbone (ablation) | Adder FCOS · COCO val2017 | mAP: NAN -> 32.4 | Helped |
| use l1 (sign) gradients for input features instead of clipped l2 (ablation) | Adder FCOS · COCO val2017 | mAP: 32.4 -> 32.6 | Helped |
| use knowledge-distillation pre-trained AdderNet backbone (ablation) | Adder FCOS · COCO val2017 | mAP: 32.6 -> 33.2 | Helped |
| add cosine learning rate and LR tuning (ablation) | Adder FCOS · COCO val2017 | mAP: 33.2 -> 34.8 | Helped |
| replace FPN with PAFPN neck (ablation) | Adder FCOS · COCO val2017 | mAP: 34.8 -> 34.9 | No clear effect |
| add shortcut connections to PAFPN (ablation) | Adder FCOS · COCO val2017 | mAP: 34.9 -> 36.1 | Helped |
| reverse fusion order (bottom-up first, R-PAFPN) (ablation) | Adder FCOS · COCO val2017 | mAP: 36.1 -> 37.0 | Helped |
| KD-pretrained instead of vanilla AdderNet-50 backbone (ablation) | FCOS with AdderNet-50 · COCO val2017 | mAP: 37.4 -> 37.8 | Helped |
| reduce batch size from 8 to 2 with unfrozen BN (ablation) | Adder FCOS · COCO val2017 | mAP: -5.4 mAP (conv: -3.4) | Hurt |
| INT8 post-training quantization | Adder FCOS (B+N) · COCO val2017 | mAP: 37.0 -> 36.2 | Hurt |
| replace backbone and neck with adder filters | FCOS (24 epochs) · COCO val2017 | mAP: 39.0 -> 37.8 | Hurt |
| cosine LR with LR tuning vs step LR (ablation) | Adder FCOS · COCO val2017 | mAP: 33.2 -> 34.8 | Helped |
Closest papers by meaning
- Exploring Salient Object Detection with Adder Neural Networks (AAAI 2025 · similarity 0.91)
- Adder Attention for Vision Transformer (NeurIPS 2021 · similarity 0.87)
- Towards Stable and Robust AdderNets (NeurIPS 2021 · similarity 0.86)
- AdderSR: Towards Energy Efficient Image Super-Resolution (CVPR 2021 · similarity 0.86)
- EfficientDet: Scalable and Efficient Object Detection (CVPR 2020 · similarity 0.86)
- Dynamic Refinement Network for Oriented and Densely Packed Object Detection (CVPR 2020 · similarity 0.84)
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.