This repository contains the implementation of the methods described in
Improving Self-Supervised Single View Depth Estimation by Masking Occlusion
We introduce an occlusion mask, that during training specifically ignores regions that cannot be reconstructed due to occlusions and prevents them from being used in the supervisory signal.
The implementation is a modified version of Monodepth2.
You can refer to the README from the original project to setup the required environment.
You can predict depth for a single image with:
python test_simple.py --image_path assets/test_image.jpg --model_name non_occluded_min_640x192On its first run this will download the non_occluded_min_640x192 pretrained model (99MB) into the models/ folder.
We provide the following options for --model_name:
--model_name |
Photometric loss | Occlusion mask | Model resolution | KITTI abs. rel. error | delta < 1.25 |
|---|---|---|---|---|---|
| Average reprojection | No | 640 x 192 | 0.117 | 0.870 | |
non_occluded_avg_640x192 |
Non-occluded avg. reprojection | Yes | 640 x 192 | 0.117 | 0.874 |
mono_640x192 |
Per-pixel min. reprojection | No | 640 x 192 | 0.115 | 0.877 |
non_occluded_min_640x192 |
Non-occluded min. reprojection | Yes | 640 x 192 | 0.113 | 0.878 |