This is code release for our paper "Incremental Learning of Object Detectors without Catastrophic Forgetting" published on ICCV 2017.
Code is written for Python 3.5 and TensorFlow 1.5 (might require minor modifications for more recent versions). You are also expected to have normal scientific stack installed: NumPy, SciPy, Matplotlib, OpenCV. If you don't like OpenCV, you can replace it with something that can read and resize images. SciPy is used only for interaction with Matlab.
You also need a checkpoint of pre-trained ResNet-50 to initialize an object detector. Put it in the directory ./resnet. These weights are obtained from official Microsoft release, but slightly changed to correspond better to TF minor differences. This checkpoint is different from the one released by Google for TF-Slim.
All experiments were done on PASCAL VOC 2007 and Microsoft COCO. To use COCO you also need pycocotools installed.
To train and evaluate a normal FastRCNN on VOC 2007 launch the following command:
python3 frcnn.py sigmoid --run_name=resnet_sigmoid_20 --num_classes=20 --dataset=voc07 --max_iterations=40000 --action=train,eval --eval_first_n=5000 --eval_ckpts=40k --learning_rate=0.001 --sigmoid
To train 10 classes network and then extend it for 10 more classes:
python3 frcnn.py sigmoid --run_name=resnet_sigmoid_10 --num_classes=10 --dataset=voc07 --max_iterations=40000 --action=train,eval --eval_ckpts=40k --learning_rate=0.001 --lr_decay 30000 --sigmoid
python3 frcnn.py sigmoid --run_name=resnet_sigmoid_10_ext10 --num_classes=10 --extend=10 --dataset=voc07 --max_iterations=40000 --action=train,eval --eval_ckpts=40k --learning_rate=0.0001 --sigmoid --pretrained_net=resnet_sigmoid_10 --distillation --bias_distillation
The same way to train a COCO model on all classes:
python3 frcnn.py --run_name=resnet_coco_80 --num_classes=80 --dataset=coco --max_iterations=500000 --lr_decay_step=250000 --weight_decay=0.00005 --eval_first_n=5000 --eval_ckpts=500000 --action=train,eval --sigmoid"