Research map / Detection and segmentation / CNN features
PointINS: Point-Based Instance Segmentation
Lu Qi, Yi Wang, Yukang Chen, Ying-Cong Chen, Xiangyu Zhang, Jian Sun, Jiaya Jia
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 |
|---|---|---|---|
| add 9 feature channels to the instance-agnostic template (shape 9x16x16 instead of 1x16x16) (ablation) | PointINS · COCO val | mask AP: 30.4 -> 32.5 | Improved |
| use 9x16x16 template reshape instead of 1x48x48 natural representation (ablation) | PointINS · COCO val | mask AP: 29.6 -> 32.5 | Improved |
| increase template channels beyond 9 (25 channels) (ablation) | PointINS · COCO val | mask AP: 32.5 -> 32.5 | No clear effect |
| add explicit box information I^z for instance-aware weights instead of detection features (ablation) | PointINS · COCO val | mask AP: 27.9 -> 32.5 | Improved |
| add the ReLU after the FC weight predictor (ablation) | PointINS · COCO val | mask AP: 31.7 -> 32.5 | Improved |
| stack a second FC+ReLU in the weight predictor (ablation) | PointINS · COCO val | mask AP: 32.5 -> 32.6 | No clear effect |
| fuse template and weights by dynamic convolution instead of addition (ablation) | PointINS · COCO val | mask AP: 31.1 -> 32.5 | Improved |
| fuse template and weights by dynamic convolution instead of concatenation (ablation) | PointINS · COCO val | mask AP: 30.9 -> 32.5 | Improved |
| use 3x3 instead of 1x1 dynamic convolution kernel (ablation) | PointINS · COCO val | mask AP: 32.5 -> 32.5 | No clear effect |
| increase sampled positive points from 32 to 128 (ablation) | PointINS · COCO val | mask AP: 29.6 -> 32.5 | Improved |
| sample 256 positive points instead of 128 (ablation) | PointINS · COCO val | mask AP: 32.5 -> 32.4 | No clear effect |
| lower anchor IoU threshold from 0.5 to 0.4 (ablation) | PointINS · COCO val | mask AP: 32.5 -> 30.2 | Hurt |
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.
- PolarMask — PolarMask: Single Shot Instance Segmentation With Polar Representation (CVPR 2020) · better on 2 of 2 shared results
- TensorMask — TensorMask: A Foundation for Dense Object Segmentation (ICCV 2019) · better on 2 of 2 shared results
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
- Pointly-Supervised Instance Segmentation (CVPR 2022 · similarity 0.92)
- Instance and Panoptic Segmentation Using Conditional Convolutions (TPAMI 2022 · similarity 0.90)
- SOLO: Segmenting Objects by Locations (ECCV 2020 · similarity 0.90)
- BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021 · similarity 0.90)
- Fully Convolutional Instance-Aware Semantic Segmentation (CVPR 2017 · similarity 0.89)
- BlendMask: Top-Down Meets Bottom-Up for Instance Segmentation (CVPR 2020 · similarity 0.89)
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.
