SfMLearner Pytorch version
This codebase implements the system described in the paper:
Unsupervised Learning of Depth and Ego-Motion from Video
In CVPR 2017 (Oral).
See the project webpage for more details.
This codebase was developed and tested with Pytorch 0.4.1, CUDA 9.1 and Ubuntu 16.04. Original code was developped in tensorflow, you can access it here
pip3 install -r requirements.txt
or install manually the following packages :
pytorch >= 0.4.1 imageio scipy argparse tensorboardX blessings progressbar2 path.py
It is also advised to have python3 bindings for opencv for tensorboard visualizations
What has been done
- Training has been tested on KITTI and CityScapes.
- Dataset preparation has been largely improved, and now stores image sequences in folders, making sure that movement is each time big enough between each frame
- That way, training is now significantly faster, running at ~0.14sec per step vs ~0.2s per steps initially (on a single GTX980Ti)
- In addition you don't need to prepare data for a particular sequence length anymore as stacking is made on the fly.
- You can still choose the former stacked frames dataset format.
- Convergence is now almost as good as original paper with same hyper parameters
- You can know compare with groud truth for your validation set. It is still possible to validate without, but you now can see that minimizing photometric error is not equivalent to optimizing depth map.
Differences with official Implementation
- Smooth Loss is different from official repo. Instead of applying it to disparity, we apply it to depth. Original disparity smooth loss did not work well (don't know why !) and it did not even converge at all with weight values used (0.5).
- loss is divided by
2.3when downscaling instead of
2. This is the results of empiric experiments, so the optimal value is clearly not carefully determined.
- As a consequence, with a smooth loss of
2.0̀, depth test is better, but Pose test is worse. To revert smooth loss back to original, you can change it here
Preparing training data
Preparation is roughly the same command as in the original code.
For KITTI, first download the dataset using this script provided on the official website, and then run the following command. The
--with-depth option will save resized copies of groundtruth to help you setting hyper parameters. The
--with-pose will dump the sequence pose in the same format as Odometry dataset (see pose evaluation)
python3 data/prepare_train_data.py /path/to/raw/kitti/dataset/ --dataset-format 'kitti' --dump-root /path/to/resulting/formatted/data/ --width 416 --height 128 --num-threads 4 [--static-frames /path/to/static_frames.txt] [--with-depth] [--with-pose]
For Cityscapes, download the following packages: 1)
camera_trainvaltest.zip. You will probably need to contact the administrators to be able to get it. Then run the following command
python3 data/prepare_train_data.py /path/to/cityscapes/dataset/ --dataset-format 'cityscapes' --dump-root /path/to/resulting/formatted/data/ --width 416 --height 171 --num-threads 4
Notice that for Cityscapes the
img_height is set to 171 because we crop out the bottom part of the image that contains the car logo, and the resulting image will have height 128.
Once the data are formatted following the above instructions, you should be able to train the model by running the following command
python3 train.py /path/to/the/formatted/data/ -b4 -m0.2 -s0.1 --epoch-size 3000 --sequence-length 3 --log-output [--with-gt]
You can then start a
tensorboard session in this folder by
and visualize the training progress by opening https://localhost:6006 on your browser. If everything is set up properly, you should start seeing reasonable depth prediction after ~30K iterations when training on KITTI.
Disparity map generation can be done with
python3 run_inference.py --pretrained /path/to/dispnet --dataset-dir /path/pictures/dir --output-dir /path/to/output/dir
Will run inference on all pictures inside
dataset-dir and save a jpg of disparity (or depth) to
output-dir for each one see script help (
-h) for more options.
Disparity evaluation is avalaible
python3 test_disp.py --pretrained-dispnet /path/to/dispnet --pretrained-posenet /path/to/posenet --dataset-dir /path/to/KITTI_raw --dataset-list /path/to/test_files_list
Test file list is available in kitti eval folder. To get fair comparison with Original paper evaluation code, don't specify a posenet. However, if you do, it will be used to solve the scale factor ambiguity, the only ground truth used to get it will be vehicle speed which is far more acceptable for real conditions quality measurement, but you will obviously get worse results.
Pose evaluation is also available on Odometry dataset. Be sure to download both color images and pose !
python3 test_pose.py /path/to/posenet --dataset-dir /path/to/KITIT_odometry --sequences 
ATE (Absolute Trajectory Error) is computed as long as RE for rotation (Rotation Error). RE between
R2 is defined as the angle of
R1*R2^-1 when converted to axis/angle. It corresponds to
RE = arccos( (trace(R1 @ R2^-1) - 1) / 2).
While ATE is often said to be enough to trajectory estimation, RE seems important here as sequences are only
seq_length frames long.
Arguments used :
python3 train.py /path/to/the/formatted/data/ -b4 -m0 -s2.0 --epoch-size 1000 --sequence-length 5 --log-output --with-gt
|Abs Rel||Sq Rel||RMSE||RMSE(log)||Acc.1||Acc.2||Acc.3|
5-frames snippets used
|Seq. 09||Seq. 10|
|ATE||0.0179 (std. 0.0110)||0.0141 (std. 0.0115)|
|RE||0.0018 (std. 0.0009)||0.0018 (std. 0.0011)|
TensorFlow by tinghuiz (original code, and paper author)