Research map / Detection and segmentation / CNN features
Efficient Segmentation: Learning Downsampling Near Semantic Boundaries
Dmitrii Marin, Zijian He, Peter Vajda, Priyam Chatterjee, Sam Tsai, Fei Yang, Yuri Boykov
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 uniform downsampling with learned boundary-driven non-uniform downsampling | segmentation network · ApolloScape val (32x32, target classes) | mIoU: 0.19 -> 0.24 | Improved |
| replace uniform downsampling with adaptive downsampling | segmentation network · ApolloScape val (512x512, target classes) | mIoU: 0.44 -> 0.47 | Improved |
| add adaptive downsampling | PSP-Net · CityScapes val (64x64) | mIoU: 0.29 -> 0.32 | Improved |
| add adaptive downsampling at higher resolution | PSP-Net · CityScapes val (256x256) | mIoU: 0.54 -> 0.54 | No clear effect |
| add adaptive downsampling | Deeplabv3+ · CityScapes val (160x160) | mIoU: 0.54 -> 0.58 | Improved |
| add adaptive downsampling | segmentation network · Synthia (target classes) | mIoU: improvement of 1.5% to 3% at negligible cost | Improved |
| add adaptive downsampling | segmentation network · Synthia (32x32, all classes) | mIoU: 0.65 -> 0.67 | Improved |
| add adaptive downsampling | segmentation network · Supervisely person (16x16) | mIoU: 0.69 -> 0.73 | Improved |
| add adaptive downsampling | segmentation network · Supervisely person (32x32) | mIoU: 0.76 -> 0.76 | No clear effect |
| focus sampling on boundaries of selected target classes | segmentation network · ApolloScape (non-target classes) | IoU per class: non-target classes sometimes lower, e.g. sidewalk 0.29 -> 0.38 but pole 0.11 -> 0.08 at 32 | Mixed |
| adaptive downsampling on small objects | segmentation network · ApolloScape recall by object size (64x64) | object recall | Improved |
| lower input resolution to 64x64 with adaptive downsampling | segmentation network · Supervisely trimap accuracy | accuracy near boundaries | Improved |
Closest papers by meaning
- Revisiting Downsampling in Semantic Segmentation: Fighting Aliasing with Dynamic Gaussian and Gabor Frequency Filters (AAAI 2026 · similarity 0.89)
- Learning to Downsample for Segmentation of Ultra-High Resolution Images (ICLR 2022 · similarity 0.88)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.87)
- When Semantic Segmentation Meets Frequency Aliasing (ICLR 2024 · similarity 0.86)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.86)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.86)
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.
