Faster ILOD: Incremental Learning for Object Detectors based on Faster RCNN
train_first_step.py: normally train the first task (standard training).
上述是我目前要做的
train_incremental.py: incrementally train the following tasks (knowledge distillation based training).
The config settings for the models and datasets are under Faster-ILOD/configs folder.
VOC dataset training
e2e_faster_rcnn_R_50_C4_1x_Source_model.yaml: config and dataset settings for source model (ResNet50) trained on VOC dataset.
e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml: config and dataset settings for target model (ResNet50) trained on VOC dataset.
The code for loading VOC dataset to the model is written on the file Faster-ILOD/maskrcnn_benchmark/data/datasets/voc.py.
- Please modify the path for putting VOC dataset on the file Faster-ILOD/maskrcnn_benchmark/config/paths_catalog.py.
- Please modify the setting for the name of old class categories (all previously trained categories) on NAME_OLD_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml.
- Please modify the setting for the name of new class categories (categories for current training task) on NAME_NEW_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml
- Please modify the setting for the name of excluded categories (categories not used, since VOC has 20 categories) on NAME_EXCLUDED_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml.
- Please modify the number of detecting categories on NUM_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Source_model.yaml (number of old categories) and the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml (number of old and new categories), repectively.