Research map / Detection and segmentation / CNN features
The Surprising Impact of Mask-Head Architecture on Novel Class Segmentation
Vighnesh Birodkar, Zhichao Lu, Siyang Li, Vivek Rathod, Jonathan Huang
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 |
|---|---|---|---|
| train mask-head with only groundtruth boxes instead of proposals+GT (ablation) | Mask R-CNN ResNet-50-FPN · COCO Non-VOC | mask mAP: 18.2 -> 25.9 | Improved |
| train mask-head with only groundtruth boxes instead of proposals+GT (ablation) | Mask R-CNN ResNet-101-FPN · COCO Non-VOC | mask mAP: 19.6 -> 29.3 | Improved |
| replace ResNet-4 mask-head with Hourglass-52 (ablation) | Deep-MARC ResNet-101-FPN · COCO Non-VOC | mask mAP: 27.4 -> 34.4 | Improved |
| replace ResNet-4 mask-head with Hourglass-52 (ablation) | Deep-MARC ResNet-101-FPN · COCO VOC (seen) | mask mAP: 40.3 -> 41.9 | No clear effect |
| replace ResNet-4 with Hourglass-52 when training with proposals+GT (ablation) | Deep-MARC · COCO Non-VOC | mask mAP: 21.0 -> 20.6 | No clear effect |
| increase Hourglass mask-head depth from 10 to 20 (ablation) | Deep-MAC Hourglass-104 · COCO Non-VOC | mask mAP: 29.4 -> 32.5 | Improved |
| replace ResNet-4 with Hourglass-52 mask-head, detection-free (ablation) | Deep-MAC · COCO Non-VOC | mIOU: 62.1 -> 79.2 | Improved |
| add long range skip connections to Hourglass-20 (ablation) | Deep-MAC HG-20 · COCO Non-VOC | mask mAP: 31.7 -> 32.2 | Improved |
| add encoder-decoder structure to Hourglass-20 (ablation) | Deep-MAC HG-20 · COCO Non-VOC | mask mAP: 29.2 -> 32.2 | Improved |
| replace 20 convs with dilated convs in ResNet-20 mask-head (ablation) | Deep-MAC ResNet-20 · COCO Non-VOC | mask mAP: 29.9 -> 30.7 | Improved |
| put hourglass depth in the mask-head instead of shared backbone (ablation) | Deep-MAC · COCO Non-VOC, GT boxes | mIOU: 68.8 -> 77.8 | Improved |
| replace Hourglass-52 mask-head with 4 fully connected layers (ablation) | Deep-MAC · COCO Non-VOC | mask mAP: 32.5 -> 28.2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ShapeMask — ShapeMask: Learning to Segment Novel Objects by Refining Shape Priors (ICCV 2019) · 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
- Prior to Segment: Foreground Cues for Weakly Annotated Classes in Partially Supervised Instance Segmentation (ICCV 2021 · similarity 0.92)
- BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021 · similarity 0.92)
- Learning to Segment Every Thing (CVPR 2018 · similarity 0.91)
- ContrastMask: Contrastive Learning To Segment Every Thing (CVPR 2022 · similarity 0.89)
- MaskBooster: End-to-End Self-Training for Sparsely Supervised Instance Segmentation (AAAI 2023 · similarity 0.89)
- MaskDiff: Modeling Mask Distribution with Diffusion Probabilistic Model for Few-Shot Instance Segmentation (AAAI 2024 · similarity 0.88)
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.
