This folder contains code to reproduce the results on end-to-end training of PointRCNN model on stereo 3D object detection.
- pytorch 1.1.0 (torchvision 0.3.0)
- torch_scatter
- opencv-python
- tqdm
- numpy
- scipy
Download PRCNN model pretrained on SDN on KITTI here and pretrained depth model here.
After downloading model weights, create depth_network/results/stack_finetune_from_sceneflow
folder and put the depth weight under the folder.
Please refer to the original PointRCNN repo for detailed install instruction.
run end2end training (on 2 TITAN RTX GPUs):
CUDA_VISIBLE_DEVICES=0,1 python train_rcnn_depth.py --gt_database "" --cfg_file cfgs/e2e.yaml \
--batch_size 4 --train_mode end2end --ckpt_save_interval 1 --ckpt <detection_model_pretrain_path> \
--epochs 10 --mgpus --finetune
CUDA_VISIBLE_DEVICES=0,1 python eval_rcnn_depth.py --cfg_file cfgs/e2e.yaml \
--depth_ckpt <end2end_trained_depth_model_path> --ckpt <end2end_trained_detector_model_path> \
--batch_size 4 --eval_mode rcnn --mgpus