Awesome

Self-Supervised Monocular Depth Hints

Jamie Watson, Michael Firman, Gabriel J. Brostow and Daniyar Turmukhambetov – ICCV 2019

We introduce Depth Hints, which improve monocular depth estimation algorithms trained from stereo pairs.

We find that photometric reprojection losses used with self-supervised learning typically have multiple local minima.
This can restrict what a regression network learns, for example causing artifacts around thin structures.

Depth Hints are complementary depth suggestions obtained from simple off-the-shelf stereo algorithms, e.g. Semi-Global Matching. These hints are used during training to guide the network to learn better weights. They require no additional data, and are assumed to be right only sometimes.

Combined with other good practices, Depth Hints gives state-of-the-art depth predictions on the KITTI benchmark (see images above and results table below). We show additional monocular depth estimation results on the sceneflow dataset:

✏️ 📄 Citation

If you find our work useful or interesting, please consider citing our paper:

@inproceedings{watson-2019-depth-hints,
  title     = {Self-Supervised Monocular Depth Hints},
  author    = {Jamie Watson and
               Michael Firman and
               Gabriel J. Brostow and
               Daniyar Turmukhambetov},
  booktitle = {The International Conference on Computer Vision (ICCV)},
  month = {October},
  year = {2019}
}

📈 KITTI Results

Model name	Training modality	ImageNet pretrained	Resolution	Abs rel	Sq rel	𝛿 < 1.25
`Ours Resnet50`	Stereo	Yes	640 x 192	0.102	0.762	0.880
`Ours Resnet50 no pt`	Stereo	No	640 x 192	0.118	0.941	0.850
`Ours HR Resnet50`	Stereo	Yes	1024 x 320	0.096	0.710	0.890
`Ours HR Resnet50 no pt`	Stereo	No	1024 x 320	0.112	0.857	0.861
`Ours HR`	Mono + Stereo	Yes	1024 x 320	0.098	0.702	0.887

Please see the paper for full results. To download the weights and predictions for each model please follow the links below:

Model name	Training modality	ImageNet pretrained	Resolution	Weights	Eigen Predictions
`Ours Resnet50`	Stereo	Yes	640 x 192	Download	Download
`Ours Resnet50 no pt`	Stereo	No	640 x 192	Download	Download
`Ours HR Resnet50`	Stereo	Yes	1024 x 320	Download	Download
`Ours HR Resnet50 no pt`	Stereo	No	1024 x 320	Download	Download
`Ours HR`	Mono + Stereo	Yes	1024 x 320	Download	Download

⚙️ Code

The code for Depth Hints builds upon monodepth2. If you have questions about running the code, please see the issues in that repository first.

To train using depth hints:

Clone this repository
Run python precompute_depth_hints.py --data_path <your_KITTI_path>, optionally setting --save_path (will default to <data_path>/depth_hints) and --filenames (will default to training and validation images for the eigen split). This will create the "fused" depth hints referenced in the paper. This process takes approximately 4 hours on a GPU.
Add the flag --use_depth_hints to your usual monodepth2 training command, optionally also setting --depth_hint_path (will default to <data_path>/depth_hints). See below for a full command.

🎉 And that's it! 🎉

👀 Reproducing Paper Results

To recreate the results from our paper, run:

python train.py
  --data_path <your_KITTI_path>
  --log_dir <your_save_path>
  --model_name stereo_depth_hints
  --use_depth_hints
  --depth_hint_path <your_depth_hint_path>
  --frame_ids 0  --use_stereo
  --scheduler_step_size 5
  --split eigen_full
  --disparity_smoothness 0

Additionally:

For Resnet50 models, add --num_layers 50
Add --height 320 --width 1024 for High Resolution models (you may also have to set --batch_size 6 depending on the size of your GPU)
For Mono + Stereo add --frame_ids 0 -1 1 and remove --split eigen_full

The results above and in the main paper arise from evaluating on the KITTI sparse LiDAR point cloud, using the Eigen Test split.

To test on KITTI, run:

python evaluate_depth.py
  --data_path <your_KITTI_path>
  --load_weights_folder <your_model_path>
  --use_stereo

Make sure you have run export_gt_depth.py to extract ground truth files.

Additionally, if you see ValueError: Object arrays cannot be loaded when allow_pickle=False, then either downgrade numpy, or change line 166 in evaluate_depth.py to

gt_depths = np.load(gt_path, fix_imports=True, encoding='latin1', allow_pickle=True)["data"]

🖼 Running on your own images

To run on your own images, run:

python test_simple.py
  --image_path <your_image_path>
  --model_path <your_model_path>
  --num_layers <18 or 50>

This will save a numpy array of depths, and a colormapped depth image.