This gives an overview of the evolution of the training process, most of those experiments didn't not give any results but they demonstrate how I progressed through this project work, until I got the final trained model in experiment pig_v4.
- duroc_6stk db, 21 images annotated, no background, no masks
- duroc_6stk db, 21 images annotated, with background, no masks
- duroc_3stk db, 22 images annotated, with background, with masks (maybe offset by 1)
- duroc_3stk db, 55 images annotated, with background, with masks (mistake -> masks are offset by 1)
- duroc_3stk db, 55 images annotated, with background, with masks
- same as v2.3 but fixed visibility and mask indexes
- Small tweaks done after v2.4, no real improvement over the result
- Same as v2.5, training resumed from checkpoint at 6k iterations
- now with 236% more images annotated! (55 -> 130)
- Notes: Training loss diverged after only 10k iterations
- another test without background and no scaling (as apparent size of the pigs does not change in the pen)
- Trained model download link: pose_iter_124000.caffemodel
- Images sequences are not ideal for traning because successive frames might be too similar from each other
- The nose is rarely visible, maybe try a new model (PIG_4) with only: LEar, REar, Neck and Tail? Or add other useful joints (shoulders, hips,...)
- Combine training data from multiples folders (3stk,6stk,10stk... duroc/landsvin...) for better generalization
- Tweak training hyperparameters
- Average Precision (& Average Recall), using Keypoint Similarity Score OKS from COCO
- Proportion of # tracked entity against the ground thruth # of entities