Finding trash in waste
summary
Claim: a camera-based top‑view system reduces subjective rejection by automatically flagging contaminants; currently 8.5% of truck loads is rejected. Pipeline uses object detection, pixel segmentation and anomaly scoring — trained with 174 labeled images, fine‑tuned YOLO + Segment Anything to extract objects, then visual and textual features (including 1,000 reference objects with cosine similarity) feed anomaly detectors. Prototype runs on four RGB cameras over a 2×2 m rack, yields a proof‑of‑principle ready for pilot‑scale deployment.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:05PMD waste collection
- 3:20National test center
- 6:14Data challenges for CV
- 7:20AI pipeline for anomalies
- 13:09Foundation models for detection
- 14:57Anomaly detection in features
- 17:27Visual and textual features
- 19:57Test results and validation
- 21:21Future steps and challenges
- 23:19Q&A: Project status
- 25:48Q&A: False negatives
- 28:38Q&A: Scaling detection
- 30:45Closing remarks
related talks
13mApplying Computer Vision to Reduce Contamination in the Recycling Stream
46mML Use Cases at Think Big Analytics
41mPractical Aspects of Applying Modern Computer Vision Technologies
29mTrashAsistant: A kivy App, which uses Deep Neural Networks...
12mCascade models in computer vision to boost accuracy and performance
41mKEYNOTE: Image processing, AI, and autonomous systems
27mComputer vision for automated quality inspection
24mAI in maritime - Autonomous ships