Deeper Exercise: Not Detecting New Classes

[henrylao]

Detector Not Detecting New Dataset Classes
The default dataset for yolov5s.pt consists of 80 classes where traffic light and stop sign from the New
Road Sign Detection Dataset. Despite training with forecasting using a validation set, why does the model NOT detect crosswalk and speed limit when deploying to Android?

• Option 1: Use the pretrained weights and train with more epochs to account for the new classes to hopefully overcome pretrained model's priors
• Option 2: Train from scratch following: https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data using --weights '' --cfg yolov5s.yaml where weights are randomly initialized
• UPDATED per (11/09) meeting with Professor Grewe -- remove the top layer so that we can reconfigure how the model activates and forecasts. NOTE since we're using PyTorch, the removal of the top-layer and reconfiguration will require modifications in the yaml file that defines a model's architecture @ManishKakarla @PradeepG00 @akailany
  - Related Issue from YoloV5: How to add another detection head ? ultralytics/yolov5#1418
  - File to Modify if we plan to use small yolo v5: https://github.com/ultralytics/yolov5/blob/199c9c787427ece5723d5309e1c7c524a99bc59d/models/yolov5s.yaml#L47