- Gradient accumulation
- Albumentations
- Leave-one-class-out cross-validation
- Better backbone
- freeze backbone for training to run faster
- crop objects from some examples and paste them to other images
- class dropout, train "any-class" model
msh - label smoothing
msh - check why new class prediction does not work: if it RPN or classification
- Reduce images, because large part of an image is not useful at all
- Try single-stage detectors
- Visualize predictions and train data (we would like to see what part of an image can be cropped out)
- TensorboardX: training progress, learning rate, val metrics, etc
- Test-time augmentation
- Normal ensembling; ensembling models from previous epochs
- Add grayscale IR images to train procedure
- Check wtf is with the books, why we do not predict them
- Generate visualizations after training
- Train on the whole train set for final submission
- Check that bbox sizes of Faster-RCNN fit our case