Course 598 deep learning project

The stable monocular depth esimation code (in .\monodepth\torch) is adpated from [https://github.com/ClubAI/MonoDepth-PyTorch.git ], which is based on https://github.com/mrharicot/monodepth.git , and was finally used for bounding box prediction. The Mask-RCNN code (in .\Mask_RCNN\) is from https://github.com/matterport/Mask_RCNN.git, and fine-tuned on KITTI by us.

git clone https://github.com/WeilinXu/DL-project.git
cd DL-project

Add pre-trained model for monodepth:

sh ./monodepth/utils/get_model.sh model_kitti ../model/monodepth_model

Add pre-trained and KITTI fine-tuned model for Mask-RCNN: Download mask_rcnn_kitti.h5 to ../model/maskecnn_model/ from the here.

Add pre-trained model for disparityToDepth network: Download trained_model_cnn.pth to ../model/d2z_model/ from here

The input left image file (eg. 000056_10.png) should be put in DL-project/images/data/left_img. The corresponding calibration file (eg. 000056.txt) should be put in DL-project/images/data/calibration. Generate depth image estimation data (eg. 000056_10_disp.npy) and depth image (eg. 000056_10_disp_pred.png) in DL-project/images/res and plot 3D bounding box:

python src/demo2.py

Alternatively, generate depth image estimation data (eg. 000056_10_disp.npy) and depth image (eg. 000056_10_disp_pred.png) in DL-project/images/res (not plot 3D bounding box):

python ./monodepth/monodepth_simple.py --image_path images/data/left_img/000056_10.png --checkpoint_path ../model/monodepth_model/model_kitti