Research map / Detection and segmentation / CNN features
Segmentation from Natural Language Expressions
Ronghang Hu, Marcus Rohrbach, Trevor Darrell
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 deconvolution upsampling and fine-tune a high resolution model over the low resolution model (ablation) | Ours · ReferIt test | precision@0.5: 29.54% -> 34.02% | Improved |
| add deconvolution upsampling and fine-tune a high resolution model over the low resolution model (ablation) | Ours · ReferIt test | precision@0.7: 13.69% -> 19.32% | Improved |
| add deconvolution upsampling and fine-tune a high resolution model over the low resolution model (ablation) | Ours · ReferIt test | precision@0.9: 0.75% -> 3.92% | Improved |
| add deconvolution upsampling and fine-tune a high resolution model over the low resolution model (ablation) | Ours · ReferIt test | overall IoU: 45.57% -> 48.03% | Improved |
| average per-word scores rather than intersect or union them (ablation) | per-word segmentation baseline · ReferIt test | overall IoU: 26.69% -> 27.23% | Improved |
| average per-word scores rather than take their union (ablation) | per-word segmentation baseline · ReferIt test | overall IoU: 19.37% -> 27.23% | Improved |
| apply GrabCut to the box instead of using the whole bounding box (ablation) | SCRC · ReferIt test | precision@0.5: 9.73% -> 11.91% | Improved |
| apply GrabCut to the box instead of using the whole bounding box (ablation) | GroundeR · ReferIt test | precision@0.5: 11.08% -> 14.09% | Improved |
| apply GrabCut to the box instead of using the whole bounding box (ablation) | SCRC · ReferIt test | overall IoU: 21.72% -> 17.84% | Hurt |
| apply GrabCut to the box instead of using the whole bounding box (ablation) | GroundeR · ReferIt test | overall IoU: 20.50% -> 20.09% | Hurt |
| use the end-to-end recurrent convolutional model instead of proposal-based GrabCut/MCG pipelines | segmentation from expressions · ReferIt test | average time per input (sec): 9.375 -> 0.325 | Improved |
| use the full model instead of the per-word segmentation baseline | segmentation from expressions · ReferIt test | average time per input (sec): 0.169 -> 0.325 | Hurt |
Closest papers by meaning
- Dynamic Multimodal Instance Segmentation Guided by Natural Language Queries (ECCV 2018 · similarity 0.90)
- See-Through-Text Grouping for Referring Image Segmentation (ICCV 2019 · similarity 0.90)
- Recurrent Multimodal Interaction for Referring Image Segmentation (ICCV 2017 · similarity 0.90)
- Referring Image Segmentation via Recurrent Refinement Networks (CVPR 2018 · similarity 0.90)
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.88)
- CLIP as RNN: Segment Countless Visual Concepts without Training Endeavor (CVPR 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.
