AnyNet

tags: AnyNet

🔰 Introduction

The goal of this repo is to re-implement the amazing work of Yan Wang et al. for Anytime Stereo Image Depth Estimation on Mobile Devices. Original code and paper could be found via the following links:

1. Original repo
2. Original paper

This repo support pytorch 1.10.0+

🌌 Difference bwtween original repo and this repo

1. build residual cost volume
   • The reason we can see this issue.
2. refine network without spn

🔑 Training

Pretrain on SceneFlow dataset

We pretrain our model on SceneFlow dataset for 10 epochs.

pretrained script

python main.py --save_path results/pretrained_anynet_refine \
               --with_refine \
               --datapath your_path

Finetune on KITTI 2015

We finetune our model on KITTI 2015 dataset for 300 epochs. Split 80% data for training and 20% for validation.

finetune KITTI 2015 script

python finetune.py --save_path results/finetune_anynet_refine \
                   --pretrained results/pretrained_anynet_refine/checkpoint.tar \
                   --with_refine \
                   --datapath your_path \
                   --datatype 2015 \
                   --split_file dataset/KITTI2015_val.txt

Finetune on KITTI 2012

We finetune our model on KITTI 2012 dataset for 300 epochs. Split 80% data for training and 20% for validation.

finetune KITTI 2012 script

python finetune.py --save_path results/finetune_anynet_refine_2012 \
                   --pretrained results/pretrained_anynet_refine/checkpoint.tar \
                   --with_refine \
                   --datapath your_path \
                   --datatype 2012 \
                   --split_file dataset/KITTI2012_val.txt

🔎 Result on KITTI 2012, 2015 Validation

Dataset	Stage 1	Stage 2	Stage 3	Stage 4
KITTI 2012	16.86%	10.68%	7.15%	7.15%
KITTI 2015	16.57%	10.58%	6.33%	6.16%

🔥 Runtime on 2080 Ti

	Stage 1	Stage 2	Stage 3	Stage 4
Runtime	5.3ms	8.45ms	11.1ms	11.4ms