Research map / Detection and segmentation / Open-set and few-shot detection
Incidents1M: A Large-Scale Dataset of Images With Natural Disasters, Damage, and Incidents
Ethan Weber, Dim P. Papadopoulos, Àgata Lapedriza, Ferda Ofli, Muhammad Imran, Antonio B. Torralba
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 class-negative labels to the training loss (ablation) | Incidents1M Model (ResNet-50) · Incidents1M test set | incident mAP: 62.83 -> 67.19 | Improved |
| add class-negative labels to the training loss (ablation) | Incidents1M Model (ResNet-50) · Incidents1M test set | place mAP: 63.15 -> 62.94 | No clear effect |
| train on 100% instead of 25% of Incidents1M (ablation) | Incidents1M Model · Incidents1M test set | incident mAP: 64.01 -> 67.19 | Improved |
| train on 100% instead of 25% of Incidents1M (ablation) | Incidents1M Model · Incidents1M test set | place mAP: 59.68 -> 62.94 | Improved |
| train on 50% instead of 25% of the data (ablation) | Incidents1M Model · Incidents1M test set | incident mAP: 64.01 -> 65.59 | Improved |
| use multi-label Incidents1M data (about 2x images) instead of the original Incidents dataset | ResNet-50 same architecture and loss · Incidents1M test set | incident mAP: 63.89 -> 67.19 | Improved |
| use multi-label Incidents1M data (about 2x images) instead of the original Incidents dataset | ResNet-50 same architecture and loss · Incidents1M test set | place mAP: 60.33 -> 62.94 | Improved |
| use only 25% of Incidents1M (446,788 images) versus original Incidents training | Incidents1M Model · Incidents1M test set | incident mAP: 63.89 vs 64.01 | No clear effect |
| filter Twitter images with the Incidents1M Model before retrieval | Twitter earthquake events · Twitter earthquake events | average precision (%): 15.87 -> 77.71 | Improved |
| filter Twitter images with the Incidents1M Model before retrieval | Twitter flood events · Twitter flood events | average precision (%): 35.68 -> 92.96 | Improved |
| evaluate on geotagged Twitter images instead of the Incidents1M test set | Incidents1M Model · Geographical bias test | mAP: 69.94 -> 53.46 (best continent EU) | Hurt |
| filter tweets by incident images for wildfire correlation with events | Incidents1M Model temporal monitoring · Twitter vs NOAA events | IoU | Mixed |
Closest papers by meaning
- MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing (NeurIPS 2025 · similarity 0.87)
- DisasterM3: A Remote Sensing Vision-Language Dataset for Disaster Damage Assessment and Response (NeurIPS 2025 · similarity 0.86)
- Flood Insights: Integrating Remote and Social Sensing Data for Flood Exposure, Damage, and Urgent Needs Mapping (AAAI 2024 · similarity 0.86)
- RSCC: A Large-Scale Remote Sensing Change Caption Dataset for Disaster Events (NeurIPS 2025 · similarity 0.85)
- DISCount: Counting in Large Image Collections with Detector-Based Importance Sampling (AAAI 2024 · similarity 0.82)
- ERF: A Benchmark Dataset for Robust Semantic Segmentation Under Extreme Rainfall Conditions (AAAI 2025 · similarity 0.82)
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.
