GroCo: Ground Constraint for Metric Self-Supervised Monocular Depth

This repository contains the inference code for the GroCo paper (ECCV24).

Example

The model can be run on an example using the notebook example.ipynb.

Weights and Benchmark

Model and annotations are provided in this google drive

Weights should be placed in a created weights folder and annotations in splits/eigen_benchmark/gt_depths.npz.

Evaluation

Evaluation on KITTI can be run using the following command:

python evaluate_depth.py

Camera Rotations can be added with the --rotations flag, with the format [Pitch, Yaw, Roll]. Rotations are limited to 5 degrees for pitch and roll and 15 degrees for yaw to not create black borders.

python evaluate_depth.py --rotations 0 0 5

Dataset preparation

Download the KITTI dataset. Data structure should be as follows:

Kitti
├── 2011_09_26
│   ├── 2011_09_26_drive_0001_sync
│   │   ├── image_02
│   │   │   └── data
│   │   ├── image_03
│   │   │   └── data
│   │   ├── oxts
│   │   │   └── data
│   │   └── velodyne_points
│   │       └── data
    ...
└── 2011_09_28
    ...
└── 2011_09_29
    ...
└── 2011_09_30
    ...
└── 2011_10_03