DeepLosslessCompression
This repo is the official pytorch implementation for End-to-end lossless compression of high precision depth maps guided by pseudo-residual (http://arxiv.org/abs/2201.03195) (accepted by Data Compression Conference 2022)
End-to-end lossless compression of high precision depth maps guided by pseudo-residual
Yuyang Wu∗ †, Wei Gao∗,†
∗SECE, Shenzhen Graduate School, Peking University, Shenzhen, China
† Peng Cheng Laboratory, Shenzhen, China
wuyy1234@stu.pku.edu.cn,gaowei262@pku.edu.cn
We provide codes relavent to the dataset DIODE as an example. To download the training dataset and validation dataset, please refer to https://diode-dataset.org/ and organize the extracted files (including train.tar.gz, val.tar.gz, data_list.zip) as follows:
file
DeepLosslessCompression
|____train
|____val
|____data_list
|____other python files
environment
Ubuntu 18.04.3 LTS
CUDA Version 10.2.89
Python 3.6.9
torchvision 0.9.1
torch 1.8.1
other packages please refer to the import part in all .py files
usage
train DIODE
python train_end2end_compression_version8_im1_9bit_im23_9bit.py --learning_rate=0.00015 --epoch_num=200 --optimizer='Adam' --scale=1 --batch_size=16 --n_channel_1=192 --n_channel_3=64 --loss_lambda=100 --loss_beta=0 --loss_gamma=0 --loss_lambda_res_esti_1=100 --loss_gamma_res_esti_1=0 --epoch_decay=20 --lr_decay=0.75
test DIODE
python inference_end2end_compression_version1_im1_9bit_im23_9bit.py --resume_1='your trained .pkl model' --crop_width=256 --crop_height=64
pretrained model can be downloaded from:
链接：https://pan.baidu.com/s/1V6PiRQuSbUK1dXtopT0-ZA
提取码：1234
