xrosliang/edgedepth-release

Github Repo for Paper "The Edge of Depth: Explicit Constraints between Segmentation and Depth"

EdgeDepth

This is the reference PyTorch implementation for training and testing depth estimation models using the method described in

The Edge of Depth: Explicit Constraints between Segmentation and Depth

Shengjie Zhu, Garrick Brazil and Xiaoming Liu

CVPR 2020

⚙️ Setup

1. Compile Morphing operaiton:

   We implement a customized Morphing Operation in our evaluation and training codes. You can still do training and evaluation without it with a sacrifice of performance. To enable it, you can do as follows:

   1. Guranttee your computer's cuda version the same as your pytorch cuda version.

   2. Type:

   cd bnmorph
   python setup.py install
   cd ..

   You should be able to successfully compile it if you can compile cuda codes in this Pytorch Tutorial

2. Prepare Kitti Data: We use Kitti Raw Dataset as well as predicted semantics label from this Paper.

   1. To download Kitti Raw Data

   wget -i splits/kitti_archives_to_download.txt -P kitti_data/

   2. Use thins Link to download precomputed semantics Label

⏳ Training

Training Code will be released soon.

📊 evaluation

1. Pretrained Model is available here

2. Precompute GroundTruth DepthMap

   python export_gt_depth.py --data_path [Your Kitti Raw Data Address] --split eigen

3. To Evaluate without using Morphing, use command:

   python evaluate_depth.py --split eigen --dataset kitti --data_path [Your Kitti Raw Data Address] --load_weights_folder [Your Model Address] --eval_stereo \
    --num_layers 50 --post_process

   To Evaluate using Morphing, use command:

   python evaluate_depth.py --split eigen --dataset kitti --data_path [Your Kitti Raw Data Address] --load_weights_folder [Your Model Address] --eval_stereo \
    --num_layers 50 --post_process --bnMorphLoss --load_semantics --seman_path [Your Predicted Semantic Label Address]

4. You should get performance similar to Entry "Ours" listed in the table:

   Method Name	Use Lidar Groundtruth?	Is morphed?	KITTI abs. rel. error	delta < 1.25
   BTS	Yes	No	0.091	0.904
   Depth Hints	No	No	0.096	0.890
   Ours	No	No	0.091	0.898
   Ours	No	Yes	0.090	0.899

Acknowledgment

Quite a few our code base come from Monodepth2 and Depth Hints