All the utility code, config and notebooks are used for the unknown object identification using the OOD framework. The following section describes the individual code.
- get_embedings.py: code to generate feature representations from pre-trained FRCNN model for given bboxes (can be ground truth or predictions themselves)
- training.py: script to fine-tune the box-head by loading the particular pre-trained models
- eval.py: Evaluation script for to generate the COCO eval metric (MAP score) for the finetuned models
- ood_distance.py: main script for Mahalanobis distance based OOD metric generation
- Notebooks:
- feature_viz_finetune-(dataset).ipynb: for experiments for e-smart dataset feature visualization and linear separability test
- ood_distance_viz.ipynb: visualizing the results of OOD detection
- Configs:
- Base-RCNN-FPN.yaml: base FRCNN-FPN model definition config
- (dataset)-trained.yaml: For loading the complete pre-trained model checkpoints, based on the training dataset
- finetune_(dataset)_trained.yaml: Config file for fine-tuning box and loading the said checkpoint