Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
User-Instructed Disparity-aware Defocus Control
Authors: Yudong Han, Yan Yang, Hao Yang, Liyuan Pan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both common datasets and the self-collected dataset demonstrate that Ui D offers superior flexibility and quality in Do F manipulation imaging. ... Extensive experiments, including a self-collected dual-pixel real-world dataset, on defocus deblurring and image refocusing, demonstrate the superior performance of our method. ... Section 4 Experiment |
| Researcher Affiliation | Academia | 1Beijing Institute of Technology, Beijing, China 2BDSI, Australian National University, Canberra, Australia 3 Yangtze Delta Region Academy of Beijing Institude of Technology, Jiaxing, China |
| Pseudocode | Yes | Table 7: Illustration of invertible block operations Inv( ) and Inv 1( ) in our method. In the forward mapping, the input and output to each block are denoted as Ul and Ul+1. During the backward mapping, only Plus (+) and Multiply ( ) needs to be inverted. ϕ1, ϕ2, ϕ3 and ϕ4 do not need to be inverted, which can be any neural networks. |
| Open Source Code | No | If the paper is accepted, the code and data will be released. |
| Open Datasets | Yes | For defocus deblurring task, we evaluate our method on widely-used DPD-blur [2] dataset and recent DP5K [27] dataset. For refocusing task, we use three datasets for evaluation, DP5K dataset, DPD-disp dataset, and our self-collected DP dataset. |
| Dataset Splits | Yes | Our model is trained using the training splits of DPD-blur and DP5K. |
| Hardware Specification | No | The text only mentions hardware used for data collection: "The DP image is captured by Canon EOS 5D Mark IV". It does not specify hardware used for running experiments or training models. |
| Software Dependencies | No | The paper mentions several tools and frameworks (Adam W optimizer, XLMRoberta Tokenizer, BEIT-3, SAM) but does not provide specific version numbers for any software dependencies, including general programming languages or libraries like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | For training, we use the Adam W optimizer [24] with β1 = 0.9, β2 = 0.999, a learning rate of 3 10 4, and a weight decay of 10 6. A cosine annealing learning rate [32] scheduler with warmup is employed, where the cycle steps, warmup steps, and minimum learning rate are set to 200, 100, and 6 10 5, respectively. For the DPD-blur dataset, the model is trained for 40k iterations with a batch size of 4. For the DP5K dataset, we train the model for 64k iterations with a batch size of 6. ... We set λ1 = 5 10 2 and λ2 = 1 10 2. Regarding other two loss supervision λcoc and λgrad, we set λcoc = 0.5 due to that a excessive large λcoc would overwhelm the useful cues in Finit learned from reblurring and deblurring task, and λgrad = 0.5 to reserve the high-frequency information and sharp the edge in restored image. |