Awesome

LiteDepth

LiteDepth: Digging into Fast and Accurate Depth Estimation on Mobile Devices

Zhenyu Li, Zehui Chen, Jialei Xu, Xianming Liu, Junjun Jiang

ECCVW 2022 (arXiv pdf)

Notice

• Redundancy version of LiteDepth. Main codes are in projects/.
• I'd like to complete the docs ASAP.

Install

This project is based on the following packages:

• python 3.7
• cuda 11.1
• cudnn 8.0.5_0
• pytorch 1.8.0

Before running, you should also install some packages facilitating training (Refer to the repos for installation details):

• mmclassification
• monocular-depth-estimation-toolbox
• pytorch-image-models
• robust_loss_pytorch

As for converting Pytorch to tfLite, you need to install:

• onnx 1.11.0
• onnx-simplifier 0.3.10
• onnx-tf 1.9.0
• tensorflow 2.5.0

That can be sort of tricky to handle the environment. I will double-check the environment to ensure its correctness.

Configs

• Basemodel config: projects/configs/configs_baseline/basemodel_crop_gradloss_vnl_robust.py
• Teacher config: projects/configs/configs_distll/swinl_w7_teacher.py
• Distill config: projects/configs/configs_distll/swinl_w7_similarity.py