Mirror3D --- init_depth_generator VNL

We've updated the dataloader and the main train/test script to support our input format (coco format).

Enforcing geometric constraints of virtual normal for depth prediction.

[NEW] Training codes have been uploaded!

This repository contains the source code of our paper: Yin Wei, Yifan Liu, Chunhua Shen, Youliang Yan, Enforcing geometric constraints of virtual normal for depth prediction (accepted for publication in ICCV' 2019).

Some Results

Framework

---

Hightlights

• State-of-the-art performance: The comparisons with other SOTA methods on NYU and KITTI are illustrated as follow. The published model trained on NYU can achieve 10.5% on absrel.

---

Installation

• Please refer to Installation.

Datasets

• NYUDV2 The details of datasets can be found here. The Eigen split of labeled images can be downloaded here. Please extract it to ./datasets. Our SOTA model is trained on the around 20K unlabled images.

• KITTI The details of KITTI benchmark for monocular depth prediction is here. We use both the official split and Eigen split. You can find the filenames here.

Model Zoo

• ResNext101_32x4d backbone, trained on NYU dataset, download here
• ResNext101_32x4d backbone, trained on KITTI dataset (Eigen split), download here
• ResNext101_32x4d backbone, trained on KITTI dataset (Official split), download here

Inference

# Run the inferece on NYUDV2 dataset
 python  ./tools/test_nyu_metric.py \
		--dataroot    ./datasets/NYUDV2 \
		--dataset     nyudv2 \
		--cfg_file     lib/configs/resnext101_32x4d_nyudv2_class \
		--load_ckpt   ./nyu_rawdata.pth 
		
# Test depth predictions on any images, please replace the data dir in test_any_images.py
 python  ./tools/test_any_images.py \
		--dataroot    ./ \
		--dataset     any \
		--cfg_file     lib/configs/resnext101_32x4d_nyudv2_class \
		--load_ckpt   ./nyu_rawdata.pth 

If you want to test the kitti dataset, please see here

Training

# Run the inferece on NYUDV2 dataset
 python  ./tools/train_nyu_metric.py \
		--dataroot    ./datasets/NYUDV2 \
		--dataset     nyudv2 \
		--cfg_file     lib/configs/resnext101_32x4d_nyudv2_class 
                #--load_ckpt *.pth
                #--resume		

Citation

@inproceedings{Yin2019enforcing,
  title={Enforcing geometric constraints of virtual normal for depth prediction},
  author={Yin, Wei and Liu, Yifan and Shen, Chunhua and Yan, Youliang},
  booktitle= {The IEEE International Conference on Computer Vision (ICCV)},
  year={2019}
}

Other works :)

We have done a new work to aim to solve the generalization problem of monocular depth prediction. Please refer to following link. :)

• Paper: DiverseDepth: Affine-invariant Depth Prediction Using Diverse Data
• Code: https://github.com/YvanYin/DiverseDepth

Contact

Wei Yin: [email protected]