Metric Self-Supervised Monocular Depth Estimation for Object Distance Estimation

Code for self-supervised monocular object depth estimation based on monodepth2. We have replaced MonoDepth2's default encoder-decoder architecture with a pre-trained relative depth estimation VIT from Depth-Anything and added the GPS loss from G2S to predict metric depth. Object detection is done by a YOLOv8 model trained on the COCO dataset.

There are many things to do, but for now, we have provided a trained model on the KITTI dataset without searching for the best training settings.

Model	Resolution	abs_rel	sq_rel	rmse	rmse_log	a1	a2	a3
G2S	1024×384	0.109	0.844	4.774	0.194	0.869	0.958	0.981
This Repo (G2S + Depth-Anything)	518×518	0.101	0.733	4.579	0.183	0.893	0.964	0.982

Inference on a custom image folder

1. Clone this repo and create the environment:

   git clone https://github.com/fardinayar/object_depth_anything.git
   conda env create -f environment.yml
   conda activate depth

2. Run

   python object_depth_estimation.py --image_folder {path to image folder} --output_type {"csv" or "video"}

   In the case of "CSV", results will be saved as a CSV file (results.csv) that lists objects in each image with their minimum depth from the camera. In the case of the "video", the same results will be saved in a video (demo.gif).

TODO

• Add code for training the depth estimation model on custom dataset

• Evaluate the depth model only in object regions

• Add more inference settings