🌱 BasicLFDisp

Official PyTorch implementation for LF Disparity Estimation of the IEEE TPAMI 2026 paper: "Diving into Epipolar Transformers for Light Field Super-Resolution and Disparity Estimation"

BasicLFDisp is a PyTorch-based open-source and easy-to-use toolbox for Light Field (LF) image Disparity Estimation. This toolbox introduces a simple pipeline to train/test your methods, and builds a benchmark to comprehensively evaluate the performance of existing methods. Our BasicLFDisp can help researchers to get access to LF image disparity estimation quickly, and facilitates the development of novel methods. Welcome to contribute your own methods to the benchmark.

✨ News & Updates

• [2026-03] 🎉 Our paper **"Diving into Epipolar Transformers for Light Field Super-Resolution and Disparity Estimation" has been accepted by *IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)!
• [2026-03] 🚀 We have released the BasicLFDisp toolbox, including the pre-trained models for our EPIT-Disp mechanism.

🧩 Contributions

• We provide a PyTorch-based open-source and easy-to-use toolbox for LF image disparity estimation.
• We re-implement a number of existing methods on the unified datasets, and develop a benchmark for performance evaluation.
• We share the codes, models, and results of existing methods to help researchers better get access to this area.

🔧 Installation

git clone https://github.com/ZhengyuLeung/BasicLFDisp.git 

📂 Dataset Preparation

We used the HCInew dataset for both training and test, and place the dataset to the folder ./datasets/.

Run Generate_Data_for_Disp_Training.py to generate training data. The generated data will be saved in ./data_for_training/.

Run Generate_Data_for_Disp_Test.py to generate test data. The generated data will be saved in ./data_for_test/.

🚀 Training

Modify the configs in train_Disp.py or use default arguments.

Checkpoints and Logs will be saved to ./log/, and the ./log/ has the following structure:

log/
├── Disp_9x9
│    ├── [model_name]_Disp
│    │     ├── [model_name]_Disp.txt    
│    │     ├── checkpoints    
│    │     │    ├── [model_name]_Disp_9x9_Disp_epoch_01_model.pth    
│    │     │    ├── [model_name]_Disp_9x9_Disp_epoch_02_model.pth    
│    │     │    └── ...    
│    │     ├── results    
│    │     │    ├── VAL_epoch_01    
│    │     │    ├── VAL_epoch_02    
│    │     │    └── ...    
│    │     ├── [other_model_name]
│    │     └── ...
│    ├── [other_model_name]
│    └── ...    
└── ...

🧪 Test

Run test_Disp.py to perform network inference.

The PSNR and SSIM values of each dataset will be saved to ./log/, and the ./log/ has the following structure:

log/
├── Disp_9x9
│    ├── [model_name]_Disp
│    │    ├── [model_name]_log.txt
│    │    ├── checkpoints    
│    │    ├── results    
│    │    │    ├── Test  
│    │    │    │    ├── evaluation_Disp_[model_name]_Disp.xlsx
│    │    │    │    ├── [dataset_name]_Disp    
│    │    │    │    │     ├── [scene_1_name]    
│    │    │    │    │     │    ├── backgammon.bmp
│    │    │    │    │     │    ├── backgammon_BP01.bmp    
│    │    │    │    │     │    ├── backgammon_BP03.bmp  
│    │    │    │    │     │    ├── backgammon_BP07.bmp  
│    │    │    │    │     │    ├── backgammon_error.bmp    
│    │    │    │    │     │    └── backgammon.pfm    
│    │    │    │    │     ├── [scene_2_name]
│    │    │    │    │     └── ...    
│    │    │    │    └── ...        
│    │    │    └── ...        
│    │    └── ...
│    ├── [other_model_name]
│    └── ...
└── ...

📊 Benchmark

We benchmark several methods on the above datasets. MSE and BP_07, BP_03, BP_01 metrics are used for quantitative evaluation. To obtain the metric score for a dataset with M scenes, we obtain the score for this dataset by averaging the scores of all its M scenes.

The definitions and meanings of the metrics are as follows:

• MSE_100: The result of multiplying Mean Squared Error (MSE) by 100. It is used to measure the average squared difference between predicted disparity and ground-truth disparity.
• BP_07: Bad Pixel Rate (BP) with an error threshold of 0.07, i.e., the percentage of pixels where the absolute error between predicted disparity and ground-truth disparity is ≥ 0.07.
• BP_03: BP with an error threshold of 0.03, i.e., the percentage of pixels where the absolute error between predicted disparity and ground-truth disparity is ≥ 0.03.
• BP_01: BP with an error threshold of 0.01, i.e., the percentage of pixels where the absolute error between predicted disparity and ground-truth disparity is ≥ 0.01.

📊 Click to expand benchmark results

Methods	#Params.	MSE_100	BP_07	BP_03	BP_01
AttMLFNet	1.547M	12.862	18.949	30.396	57.802
DistgDisp	7.982M	2.412	7.379	14.506	37.192
EPI_ORM	5.097M	11.573	34.763	64.211	87.445
Epinet	11.02M	8.843	44.867	69.069	88.661
LFattNet	1.753M	27.162	30.622	49.776	78.612
OACC_Net	5.018M	4.495	12.124	20.091	46.594
EPIT_Disp	1.206M	3.802	15.463	28.266	59.825
EPIT_Disp_C128	1.206M	4.335	15.985	28.776	58.415

⬇️ Recources

• The pre-trained models of the aforementioned methods can be downloaded via this link.

📝 Citation

If you find this code or our paper useful for your research, please consider citing:

@article{EPIT2026,
 title = {Diving into Epipolar Transformers for Light Field Super-Resolution and Disparity Estimation},
 author = {Liang, Zhengyu and Wang, Yingqian and Wang, Longguang and Yang, Jungang and Guo, Yulan and Liu, Li and Zhou, Shilin and An, Wei},
 journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
 year = {2026},
}