Research map / Detection and segmentation / CNN features
Cars Can't Fly Up in the Sky: Improving Urban-Scene Segmentation via Height-Driven Attention Networks
Check 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 height-driven attention to DeepLabv3+ | Cityscapes segmentation models with ShuffleNetV2, MobileNetV2, ResNet-50, or ResNet-101 backbones · Cityscapes validation | mIoU: 70.27 -> 71.30; 70.85 -> 71.52; 73.93 -> 74.96; 73.40 -> 74.70; 76.84 -> 77.78; 77.76 -> 78.71; 77.80 -> 79.31; 79.25 -> | Helped |
| Add height-driven attention | BDD100K segmentation models with MobileNetV2 or ResNet-101 · BDD100K validation | mIoU: 58.91% -> 60.05%; 63.75% -> 64.56%; 64.84% -> 65.60% | Helped |
| Add HANet at layers 1–4 with positional encoding (ablation) | ResNet-101 DeepLabv3+ · Cityscapes validation | mIoU: 77.80 -> 79.31 | Helped |
| Add HANet at layer 5 in addition to layers 1–4 (ablation) | ResNet-101 DeepLabv3+ with HANet at layers 1–4 · Cityscapes validation | mIoU: 79.31 -> 78.79 | Hurt |
| Change channel reduction ratio from 32 to 16 (ablation) | HANet at layers 1–4 · Cityscapes validation | mIoU: 79.31 -> 79.15 | Hurt |
| Change channel reduction ratio from 32 to 64 (ablation) | HANet at layers 1–4 · Cityscapes validation | mIoU: 79.31 -> 79.08 | Hurt |
| Remove positional encoding (ablation) | HANet at layers 1–4 · Cityscapes validation | mIoU: 79.31 -> 78.25 | Hurt |
| Replace average pooling with max pooling (ablation) | HANet at layers 1–4 · Cityscapes validation | mIoU: 79.31 -> 78.87 | Hurt |
| Replace sinusoidal encoding with learnable embedding from scratch at the second HANet convolution (ablation) | HANet · Cityscapes validation | mIoU: 80.29 -> 79.60 | Hurt |
| Replace sinusoidal encoding with pretrained learned embedding at the second HANet convolution (ablation) | HANet · Cityscapes validation | mIoU: 80.29 -> 79.30 | Hurt |
| Move sinusoidal positional encoding from the second to the first HANet convolution (ablation) | HANet · Cityscapes validation | mIoU: 80.29 -> 79.61 | Hurt |
| Use width-wise rather than height-wise attention (ablation) | Urban-scene segmentation attention network · Urban-scene segmentation | mIoU | No clear effect |
Closest papers by meaning
- Dual Attention Network for Scene Segmentation (CVPR 2019 · similarity 0.89)
- AttaNet: Attention-Augmented Network for Fast and Accurate Scene Parsing (AAAI 2021 · similarity 0.87)
- Attention-Guided Unified Network for Panoptic Segmentation (CVPR 2019 · similarity 0.87)
- Seamless Scene Segmentation (CVPR 2019 · similarity 0.87)
- RANet: Region Attention Network for Semantic Segmentation (NeurIPS 2020 · similarity 0.87)
- Learning Content-Enhanced Mask Transformer for Domain Generalized Urban-Scene Segmentation (AAAI 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.