Research map / Detection and segmentation / Open-set and few-shot detection
Traffic Scene Parsing through the TSP6K Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TSP6K, a 6,000-image pixel- and instance-labeled traffic monitoring dataset, benchmarks scene parsing, instance segmentation and UDA methods on it, and proposes a detail refining decoder on SegNeXt that improves mIoU/iIoU.
Problem
Models trained on driving-scene traffic datasets perform poorly on traffic monitoring scenes (high-mounted cameras, crowded, wide instance-size variance), and no class-labeled monitoring dataset existed for semantic and instance segmentation.
Approach
A new class-labeled traffic monitoring dataset with semantic and instance masks, plus a detail refining decoder that uses learnable region tokens with cross-attention over high-resolution features to refine per-region pixel relationships on top of the SegNeXt encoder.
Key results
- The proposed detail refining decoder on the SegNeXt-Base (MSCAN-Base) encoder reaches 75.8% mIoU and 58.4% iIoU on the TSP6K validation set, 1.2% and 1.1% above SegNeXt-Base.
- Best UDA method HRDA reaches 54.1% mIoU from Cityscapes to TSP6K, far below the fully supervised 72.4%.
- QueryInst achieves the best instance segmentation mask AP of 31.5 on validation, with models trained on TSP6K over 10% AP below Mask-RCNN's Cityscapes result.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the detail refining decoder to SegNeXt | SegNeXt with MSCAN-Base · TSP6K validation | mIoU: 74.6 -> 75.8 | Helped |
| Add the detail refining decoder to SegNeXt | SegNeXt with MSCAN-Base · TSP6K validation | iIoU: 57.3 -> 58.4 | Helped |
| Use HRDA for unsupervised domain adaptation | DeepLab-v2 source-domain baseline · SYNTHIA → TSP6K | mIoU: 21.7 -> 45.4 | Helped |
| Use HRDA for unsupervised domain adaptation | DeepLab-v2 source-domain baseline · Cityscapes → TSP6K | mIoU: 26.1 -> 54.1 | Helped |
| Use Cityscapes rather than SYNTHIA as the source domain with HRDA | HRDA · TSP6K | mIoU: 45.4 -> 54.1 | Helped |
Surprising findings
- EncNet, a self-attention method, does not improve over FCN on TSP6K, attributed to channel-wise attention losing local detail.
- Pure CNN and lightweight transformer models such as SegNeXt outperform larger Swin-based models at a fraction of the parameters and FLOPs.
Limitations
- Dataset has only 6,000 labeled images with many unlabeled images remaining.
- Geographically limited, lacking scenes from left-hand driving countries.
- RGB-only modality, limiting multi-modal model development.
Experimental setup
Datasets: TSP6K SYNTHIA Cityscapes
Benchmarks: TSP6K
Metrics: mIoU iIoU AP box AP mask
Baselines: FCN PSPNet DeepLabv3 UperNet DeepLabv3+ SegFormer Swin-UperNet SegNeXt SETR Mask-RCNN QueryInst Mask2Former
Builds on: SegNeXt DeepLabv3+ Mask2Former HRDA
Closest papers by meaning
- Weakly Supervised Semantic Segmentation for Driving Scenes (AAAI 2024 · similarity 0.86)
- MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding (CVPR 2024 · similarity 0.85)
- TUMTraf VideoQA: Dataset and Benchmark for Unified Spatio-Temporal Video Understanding in Traffic Scenes (ICML 2025 · similarity 0.85)
- eTraM: Event-based Traffic Monitoring Dataset (CVPR 2024 · similarity 0.84)
- Vanishing-Point-Guided Video Semantic Segmentation of Driving Scenes (CVPR 2024 · similarity 0.84)
- On-Road Object Importance Estimation: A New Dataset and A Model with Multi-Fold Top-Down Guidance (NeurIPS 2024 · similarity 0.84)
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.