Reproduce “StreetView-Waste: A Multi-Task Dataset for Urban Waste Management” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Diogo J. Paulo, João Martins, Hugo Proença, João C. Neves
StreetView-Waste introduces a multi-task dataset for urban waste management, comprising 36,478 images captured by vehicle-mounted fisheye cameras with annotations for container detection, multi-object tracking, and waste overflow segmentation. The work establishes baselines using current computer vision models and proposes strategies that reveal challenges in maintaining object identities and effectively integrating geometric cues for amorphous waste targets.
Key insights
- State-of-the-art multi-object trackers struggle with maintaining consistent identities for waste containers in dynamic urban scenes with frequent occlusions, as shown by the trade-off between counting accuracy and identity preservation when applying simple heuristics.
- Geometric priors (depth, surface normals) can enhance overflow segmentation, particularly for lightweight models (e.g., YOLACT), by aiding in distinguishing amorphous waste from background elements, while more complex architectures (e.g., Mask R-CNN) may not universally benefit or even degrade.
- The dataset acts as a diagnostic tool, revealing architectural limitations of current models in handling unstructured targets like overflowing waste, especially when integrating multimodal inputs or dealing with long-tail class distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.