# AECR-Net1

**Repository Path**: damon_one/AECR-Net1

## Basic Information

- **Project Name**: AECR-Net1
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2024-03-12
- **Last Updated**: 2024-03-12

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Repository to host customized AECR-Net model developed as part of [5th UG2+ challenge (CVPR 2022) Track 1.1](https://codalab.lisn.upsaclay.fr/competitions/1235)

## Paper
[Contrastive Learning for Compact Single Image Dehazing, CVPR2021](https://arxiv.org/abs/2104.09367)

## Summary

We modify the [official implementation of AECR-Net](https://github.com/GlassyWu/AECR-Net) to use it's NH_train pretrained model to perform dehazing and provide output images which serves as the input for the object detection task downstream.


We fine tuned the [NH_train  pretrainedmodel](https://github.com/ma08/AECR-Net/blob/6c604b1570e3018ab0f6723e5e4757b404f04d34/trained_models/NH_train.pk) by training it on a subset of the training set available for the competition and obtained [the best model](https://github.com/ma08/AECR-Net/blob/6c604b1570e3018ab0f6723e5e4757b404f04d34/trained_models/NH_train.pk.best) based on evaluation of the remaining subset of the training set.

The results on the finetuned model are not satisfactory when compared to DW-GAN as the latter was found to perform better for the dehazing task for the competition. It is unclear during the time of submission if tinkering with the finetuning would improve the performance.

### Testset Metrics
Before finetuning: `SSIM: 0.6288 PSNR:14.2729`
After finetuning, with the best model: `SSIM:0.8001 PSNR:20.0715`

### Sample Input and Output

Input 
![Input](NH_test/075.jpg)
Output
![Output](NH_test/output/pred.jpg)