Research map / Detection and segmentation / CNN features
CARAFE: Content-Aware ReAssembly of FEatures
Jiaqi Wang, Kai Chen, Rui Xu, Ziwei Liu, Chen Change Loy, Dahua Lin
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 |
|---|---|---|---|
| replace nearest upsampling in FPN with CARAFE | Faster R-CNN R50 · COCO 2017 val | box AP: 36.5 -> 37.8 | Improved |
| add an extra 3x3 conv after nearest or bilinear upsampling | Faster R-CNN FPN · COCO 2017 val | box AP: 36.5 -> 36.6 | No clear effect |
| replace nearest upsampling in FPN with CARAFE | Faster R-CNN R50 · COCO test-dev | box AP: 36.9 -> 38.1 | Improved |
| replace nearest in FPN and deconv in mask head with CARAFE | Mask R-CNN R50 · COCO test-dev | mask AP: 34.6 -> 35.9 | Improved |
| replace deconvolution in the mask head with CARAFE | Mask R-CNN R50 · COCO 2017 val | mask AP: 34.2 -> 34.7 | Improved |
| use CARAFE in both FPN and mask head instead of neither | Mask R-CNN R50 · COCO 2017 val | mask AP: 34.2 -> 35.7 | Improved |
| replace upsamplers in UperNet with CARAFE | UperNet R50 · ADE20k val | mIoU: 40.44 -> 42.23 | Improved |
| use CARAFE in all of PPM, FPN and FUSE rather than PPM only (ablation) | UperNet R50 · ADE20k val | mIoU: 40.85 -> 42.23 | Improved |
| replace upsampling with CARAFE in Global&Local inpainting | Global&Local · Places val | PSNR (dB): 19.58 -> 20.71 | Improved |
| replace upsampling with CARAFE in Partial Conv inpainting | Partial Conv · Places val | PSNR (dB): 20.78 -> 20.98 | Improved |
| compress channels to Cm=64 or remove the compressor (ablation) | CARAFE · COCO 2017 val | box AP: 37.8 -> 37.8 | No clear effect |
| reduce compressed channels Cm from 64 to 16 (ablation) | CARAFE · COCO 2017 val | box AP: 37.8 -> 37.6 | Hurt |
Later papers that compared with this method
8 later accepted papers report CARAFE as a baseline in their results tables; 8 of them report a better number on most of the results they share with it.
- FeatUp: A Model-Agnostic Framework for Features at Any Resolution (ICLR 2024 · better on 4 of 4 shared results)
- FADE: A Task-Agnostic Upsampling Operator for Encoder–Decoder Architectures (IJCV 2024 · better on 2 of 3 shared results)
- Learning to Upsample by Learning to Sample (ICCV 2023 · better on 4 of 4 shared results)
- SAPA: Similarity-Aware Point Affiliation for Feature Upsampling (NeurIPS 2022 · better on 3 of 4 shared results)
- FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling (ECCV 2022 · better on 4 of 4 shared results)
- Learning Implicit Feature Alignment Function for Semantic Segmentation (ECCV 2022 · better on 1 of 1 shared result)
- Dynamic Neural Representational Decoders for High-Resolution Semantic Segmentation (NeurIPS 2021 · better on 1 of 1 shared result)
- A Holistically-Guided Decoder for Deep Representation Learning With Applications to Semantic Segmentation and Object Detection (TPAMI 2021 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- PSANet — PSANet: Point-wise Spatial Attention Network for Scene Parsing (ECCV 2018) · better on 1 of 1 shared result
- Faster R-CNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CARAFE++: Unified Content-Aware ReAssembly of FEatures (TPAMI 2021 · similarity 0.98)
- Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling (CVPR 2026 · similarity 0.85)
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (TPAMI 2017 · similarity 0.85)
- TAFE-Net: Task-Aware Feature Embeddings for Low Shot Learning (CVPR 2019 · similarity 0.85)
- Lightweight and Progressively-Scalable Networks for Semantic Segmentation (IJCV 2023 · similarity 0.85)
- FADE: A Task-Agnostic Upsampling Operator for Encoder–Decoder Architectures (IJCV 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.
