Research map / Detection and segmentation / CNN features
Referring Image Segmentation via Recurrent Refinement Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 recurrent refinement with convolutional LSTM and multi-scale features | Referring image segmentation with DeepLab ResNet-101 · UNC testA | overall IoU: 45.69 -> 57.26 | Helped |
| Add recurrent refinement with convolutional LSTM and multi-scale features | Referring image segmentation with DeepLab ResNet-101 · ReferIt test | overall IoU: 58.73 -> 63.63 | Helped |
| Use C5, C4, and C3 features rather than C5 alone (ablation) | Plain RRN · UNC val | overall IoU: 46.87 -> 49.74 | Helped |
| Add lower-level C2 and C1 features to the C5, C4, C3 pyramid (ablation) | Plain RRN · UNC val | overall IoU: 49.74 -> 48.30 | Hurt |
| Replace the recurrent update with FPN-style element-wise addition (ablation) | RRN with C5, C4, and C3 features · UNC val | overall IoU: 49.74 -> 39.31 | Hurt |
| Share update weights across feature levels using a vanilla RNN (ablation) | Plain RRN with C5, C4, and C3 features · UNC val | overall IoU: 49.74 -> 48.86 | Hurt |
| Use convolutional LSTM gates rather than a vanilla RNN update (ablation) | RRN with C5, C4, and C3 features · UNC val | overall IoU: 48.86 -> 52.48 | Helped |
| Use tanh rather than ReLU after convolutional LSTM updates (ablation) | RRN with C5, C4, and C3 features · UNC val | overall IoU: 52.48 -> 54.26 | Helped |
| Use tanh rather than ReLU after vanilla RNN updates (ablation) | RRN with C5, C4, and C3 features · UNC val | overall IoU: 48.86 -> 46.35 | Hurt |
| Use distinct pyramid features instead of repeating C5 at every step (ablation) | Convolutional LSTM RRN · UNC val | overall IoU: 51.41 -> 54.26 | Helped |
| Use C5, C4, C3 feature order rather than inverse C3, C4, C5 order (ablation) | Convolutional LSTM RRN with tanh · UNC val | overall IoU: 53.81 -> 54.26 | Helped |
| Add DenseCRF post-processing (ablation) | RRN with convolutional LSTM · UNC+ testA | overall IoU: 41.68 -> 42.15 | Helped |
Closest papers by meaning
- Semantics-Aware Dynamic Localization and Refinement for Referring Image Segmentation (AAAI 2023 · similarity 0.89)
- CLIP as RNN: Segment Countless Visual Concepts without Training Endeavor (CVPR 2024 · similarity 0.88)
- ReSTR: Convolution-Free Referring Image Segmentation Using Transformers (CVPR 2022 · similarity 0.87)
- Recurrent Scene Parsing With Perspective Understanding in the Loop (CVPR 2018 · similarity 0.87)
- End-To-End Instance Segmentation With Recurrent Attention (CVPR 2017 · similarity 0.87)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 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.