Implementation for paper "STAR: A Structure-aware Lightweight Transformer for Real-time Image Enhancement" (ICCV 2021).

Last update: Dec 21, 2022

Related tags

Deep Learning STAR-pytorch

Overview

STAR-pytorch

Implementation for paper "STAR: A Structure-aware Lightweight Transformer for Real-time Image Enhancement" (ICCV 2021).

CVF (pdf)

STAR-DCE

The pytorch implementation of low light enhancement with STAR on Adobe-MIT FiveK dataset. You can find it in STAR-DCE directory. Here we adopt the pipleline of Zero-DCE ( paper | code ), just replacing the CNN backbone with STAR. In Zero-DCE, for each image the network will regress a group of curves, which will then applied on the source image iteratively. You can find more details in the original repo Zero-DCE.

Requirements

numpy
einops
torch
torchvision
opencv

Datesets

We provide download links for Adobe-MIT FiveK datasets we used ( train | test ). Please note that we adopt the test set splited by DeepUPE for fair comparison.

Training DCE models

To train a original STAR-DCE model,

cd STAR-DCE
python train_dce.py 
  --lowlight_images_path "dir-to-your-training-set" \
  --parallel True \
  --snapshots_folder snapshots/STAR-ori \
  --lr 0.001 \
  --num_epochs 100 \
  --lr_type cos \
  --train_batch_size 32 \
  --model STAR-DCE-Ori \
  --snapshot_iter 10 \
  --num_workers 32 \

To train the baseline CNN-based DCE-Net (w\ or w\o Pooling),

cd STAR-DCE
python train_dce.py 
  --lowlight_images_path "dir-to-your-training-set" \
  --parallel True \
  --snapshots_folder snapshots/DCE \
  --lr 0.001 \
  --num_epochs 100 \
  --lr_type cos \
  --train_batch_size 32 \
  --model DCE-Net \
  --snapshot_iter 10 \
  --num_workers 32 \

cd STAR-DCE
python train_dce.py 
  --lowlight_images_path "dir-to-your-training-set" \
  --parallel True \
  --snapshots_folder snapshots/DCE-Pool \
  --lr 0.001 \
  --num_epochs 100 \
  --lr_type cos \
  --train_batch_size 32 \
  --model DCE-Net-Pool \
  --snapshot_iter 10 \
  --num_workers 32 \

Evaluation of trained models

To evaluated the STAR-DCE model you trained,

cd STAR-DCE
  python test_dce.py \
  --lowlight_images_path  "dir-to-your-test-set" \
  --parallel True \
  --snapshots_folder snapshots_test/STAR-DCE \
  --val_batch_size 1 \
  --pretrain_dir snapshots/STAR-ori/Epoch_best.pth \
  --model STAR-DCE-Ori \

To evaluated the DCE-Net model you trained,

cd STAR-DCE
  python test_dce.py \
  --lowlight_images_path  "dir-to-your-test-set" \
  --parallel True \
  --snapshots_folder snapshots_test/DCE \
  --val_batch_size 1 \
  --pretrain_dir snapshots/DCE/Epoch_best.pth \
  --model DCE-Net \

Citation

If this code helps your research, please cite our paper :)

@inproceedings{zhang2021star,
  title={STAR: A Structure-Aware Lightweight Transformer for Real-Time Image Enhancement},
  author={Zhang, Zhaoyang and Jiang, Yitong and Jiang, Jun and Wang, Xiaogang and Luo, Ping and Gu, Jinwei},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={4106--4115},
  year={2021}
}

Implementation for paper "STAR: A Structure-aware Lightweight Transformer for Real-time Image Enhancement" (ICCV 2021).

Related tags

Overview

STAR-pytorch

STAR-DCE

Requirements

Datesets

Training DCE models

Evaluation of trained models

Citation

Owner

code for our paper "Source Data-absent Unsupervised Domain Adaptation through Hypothesis Transfer and Labeling Transfer"

Tutorial on active learning with the Nvidia Transfer Learning Toolkit (TLT).

Pytorch reimplement of the paper "A Novel Cascade Binary Tagging Framework for Relational Triple Extraction" ACL2020. The original code is written in keras.

NEATEST: Evolving Neural Networks Through Augmenting Topologies with Evolution Strategy Training

A DNN inference latency prediction toolkit for accurately modeling and predicting the latency on diverse edge devices.

VIMPAC: Video Pre-Training via Masked Token Prediction and Contrastive Learning

Code for the ICASSP-2021 paper: Continuous Speech Separation with Conformer.

Jittor 64*64 implementation of StyleGAN

Dynamic Token Normalization Improves Vision Transformers

ZeroVL - The official implementation of ZeroVL

Mosaic of Object-centric Images as Scene-centric Images (MosaicOS) for long-tailed object detection and instance segmentation.

ByteTrack with ReID module following the paradigm of FairMOT, tracking strategy is borrowed from FairMOT/JDE.

Learning from Synthetic Data with Fine-grained Attributes for Person Re-Identification

Using LSTM write Tang poetry

Gradient representations in ReLU networks as similarity functions

Multi-Joint dynamics with Contact. A general purpose physics simulator.

TensorFlow Ranking is a library for Learning-to-Rank (LTR) techniques on the TensorFlow platform

v objective diffusion inference code for JAX.

An implementation for the ICCV 2021 paper Deep Permutation Equivariant Structure from Motion.

A weakly-supervised scene graph generation codebase. The implementation of our CVPR2021 paper ``Linguistic Structures as Weak Supervision for Visual Scene Graph Generation''