PointRCNN-PL-end2end

This folder contains code to reproduce the results on end-to-end training of PointRCNN model on stereo 3D object detection.

Dependency

• pytorch 1.1.0 (torchvision 0.3.0)
• torch_scatter
• opencv-python
• tqdm
• numpy
• scipy

Pre-train models

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.

Install

Please refer to the original PointRCNN repo for detailed install instruction.

Run PointRCNN_PL_end2end training and evaluation

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

Evaluation

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