Research map / Detection and segmentation / Open-set and few-shot detection

UrbanWaste: In-the-Bin Dataset for Waste Disposal Inspection with Multi-Granularity Hierarchical Labels

AAAI 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Remote sensing segmentation and change · 1 citations (OpenAlex)

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 UrbanWaste, a 25,254-image in-the-bin waste segmentation/detection dataset with 193 hierarchical categories, benchmarks standard detectors and segmenters on it, and describes a deployed change-detection plus segmentation waste-delivery inspection workflow.

Problem

Residents often misclassify household waste when disposing of it, and existing waste datasets are small, clean-background, or limited to recyclables, so they cannot support in-bin supervision of garbage delivery.

Approach

Capture images from inside residential trash bins before and after each delivery, annotate dense instance masks with a four-level coarse-to-fine label hierarchy, and combine change detection with semantic segmentation to score whether a delivery was correctly sorted.

Key results

• Cascade R-CNN reaches the best detection AP of 36.56 on UrbanWaste among the compared detectors, and DeepLabv3+ reaches 60.01 mIoU and 52.85 pixel accuracy.
• In cross-dataset tests, a DeepLabv3+ model fine-tuned on UrbanWaste reaches 91.54 mIoU on TACO and 90.31 on ZeroWaste, versus 13.26 and 12.53 for models trained on TACO and ZeroWaste and tested on UrbanWaste.
• The dataset has 25,254 images and 140,008 annotated instances across 193 categories, averaging 5.54 instances per image.

Measured effects reported in the paper

Surprising findings

• Models fine-tuned on TACO and ZeroWaste transfer poorly to UrbanWaste (mIoU drops to about 13), while the UrbanWaste-trained model generalizes well to them.

Limitations

• Class distribution is highly imbalanced, which may hurt recognition of rare categories.
• Hazardous-waste bins produced little to no images, so that category is largely absent.
• Most baselines are only reported on one granularity level (the second), and the detailed experiments are deferred to supplementary material.

Experimental setup

Datasets: UrbanWaste TACO ZeroWaste

Benchmarks: UrbanWaste TACO ZeroWaste

Metrics: AP AP50 AP75 mIoU pixel accuracy

Baselines: Faster R-CNN Mask R-CNN Cascade R-CNN DeepLabv3+ YOLOv5 RetinaNet TridentNet YOLOv11

Builds on: Mask R-CNN DeepLabv3+ ChangeFormer SAM2

Closest papers by meaning

• WISNet: Pseudo Label Generation on Unbalanced and Patch Annotated Waste Images (CVPR 2025 · similarity 0.88)
• UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.81)
• Infusing Synthetic Data with Real-World Patterns for Zero-Shot Material State Segmentation (NeurIPS 2024 · similarity 0.79)
• UrbanVLP: Multi-Granularity Vision-Language Pretraining for Urban Socioeconomic Indicator Prediction (AAAI 2025 · similarity 0.78)
• SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification (NeurIPS 2024 · similarity 0.78)
• CityLens: Evaluating Large Vision-Language Models for Urban Socioeconomic Sensing (ICLR 2026 · similarity 0.78)