This package provides a python toolkit for evaluation on the SeasonDepth dataset(test set download, validation set download, training set download, paper). SeasonDepth is a new monocular depth prediction dataset with multi-traverse changing environments. Several representative baseline methods with recent open-source state-of-the-art models on KITTI leaderboard[1] are evaluated on the SeasonDepth dataset. Check more on SeasonDepth Prediction Challenge.

pip install opencv-python
pip install xlwt xlrd tqdm
pip install numpy

Download SeasonDepth dataset and create the directories below. The slices in detail can be found here.

└── SeasonDepth_DATASET_ROOT
    ├── depth
    │   ├── slice2
    │   │   ├──  img_00119_c0_1303398474779439us.png
    │   │   ├──  *.png
    │   ├── slice3
    │   ├── slice7
    │   └── slice8
    ├── images
    │   ├── slice2
    │   │   ├──  img_00119_c0_1303398474779439us.jpg
    │   │   ├──  *.jpg
    │   ├── slice3
    │   ├── slice7
    │   └── slice8

Evaluate your model on SeasonDepth and organize your results as follows:

└── YOUR_RESULT_ROOT
    ├── slice2
    │   ├──  img_00119_c0_1303398474779439us.png
    │   ├──  *.png
    ├── slice3
    ├── slice7
    └── slice8

After that, run the evaluation script in the folder eval_code to find your evaluation results.

python evaluation.py --pred_pth YOUR_RESULT_ROOT --gt_pth SeasonDepth_DATASET_ROOT

You can also add some following arguments if you want:

--gui # To watch the results of evaluation. Press q to exit and any other key to continue.
--disp2depth # To convert disparity map to depth map for correct evaluation.
--not_clean # To generate all the intermediate xls files during evaluating.

We use the following method to evaluate rank in SeasonDepth Challenge.

final_score = 0.5 * (RANK(absrel_avg) + RANK(a1_avg)) + 0.4 * (RANK(absrel_var) + RANK(a1_var)) + 0.1 * (RANK(absrel_rng) + RANK(a1_rng))

Final score of a method is related to its rank of all metrics, so your total score may change before deadline.

Some baselines on SeasonDepth test set are evaluated below. The full leadboard of the test set can be found here with more models.

Also, we present the cross-dataset performance of monocular depth estimation from KITTI to the SeasonDepth dataset over our validation set. Please refer to our paper for more details.

Qualitative comparison results with illumination or vegetation changes are shown below and more can be found in our paper.

Please cite the following papers if you use our dataset:

@article{SeasonDepth2023hu,
  title={SeasonDepth: Cross-Season Monocular Depth Prediction Dataset and Benchmark under Multiple Environments},
  author={Hu, Hanjiang and Yang, Baoquan and Qiao, Zhijian and Liu, Shiqi and Zhu, Jiacheng and Liu, Zuxin and Ding, Wenhao and Zhao, Ding and Wang, Hesheng},
  journal={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2023},
  organization={IEEE}
}

[1] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, "Vision meets robotics: The KITTI dataset," Int. J. Robot. Research (IJRR), vol. 32, no. 11, pp. 1231–1237, Sep. 2013. http://www.cvlibs.net/datasets/kitti/