AI City Challenge - Key Lessons
Key Notes
- Detection Models - YOLO2, DenseNet,Mask
- R-CNN, Faster R-CNN
- The Mask R-CNN model, in particular, was
- able to detect and localize small vehicles with excellent precision
- Tracking - clustering-based association, graph optimization, medianflow, Kalman filtering
- Most successful approaches are based on traffic motion flow analysis (e.g., using optical flow) rather than trying to detect and track individual vehicles
- Re-identification matching - triplet loss
- Re-id using Vehicle number match
- Data Issues - video quality, illumination and environmental conditions
Key Notes
- Extracting visual features from convolutional neural networks (CNNs), and leveraging semantic features from traveling direction and vehicle type classification.
- The utilization of vehicles semantic attributes
- Novel two-stage framework based on anomaly candidate identification and starting time estimation
- Data Issues - vehicle-based problems are more challenging, due to the high intra-class variability caused by the dependence of shapes on viewing angles, and high inter-class similarity, as vehicle models produced by different manufacturers look visually alike
Key Notes
- For vehicle detection, most teams [5, 1, 37] selected YOLOv3
- For vehicle tracking, DeepSORT
- Hungarian matching algorithm to associate detections into tracklets, considering both spatial and appearance features.
- Motion-based tracking
- Trained classifiers for vehicle type, color, and viewpoint/orientation using the labels on synthetic data and made predictions on real-world data
- Multi-target single-camera (MTSC) tracking,
- ReID for appearance feature extraction, and spatio-temporal association to assign identities to tracklets across multiple cameras.
- Anomaly prediction module used K-means clustering to identify potential anomalous regions
- Data Challenges Challenges in this regard include the variety in camera views, image quality, lighting, and weather conditions
Keep Thinking!!!
No comments:
Post a Comment