PhoCoLens: Photorealistic and Consistent Reconstruction in Lensless Imaging

Authors: Xin Cai, Zhiyuan You, Hailong Zhang, Jinwei Gu, Wentao Liu, Tianfan Xue

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare our approach with other methods and conduct a comprehensive ablation study to evaluate the effectiveness of our design. ... We assess the performance of our proposed method by comparing it against both traditional and learning-based approaches, using qualitative and quantitative measures on the two datasets.
Researcher Affiliation Collaboration Xin Cai1,2, Zhiyuan You1, Hailong Zhang3, Wentao Liu2,4, Jinwei Gu1, Tianfan Xue1,2 1The Chinese University of Hong Kong, 2Shanghai Artificial Intelligence Laboratory, 3Tsinghua University, 4Sense Time
Pseudocode No The paper describes the system and methods in text and diagrams (e.g., Fig. 5) but does not include formal pseudocode or algorithm blocks.
Open Source Code Yes Project website: phocolens.github.io. ... Code and data will be available on the project website.
Open Datasets Yes The Phlat Cam dataset [17] contains 10,000 images across 1,000 classes resized to 384 384 pixels. ... The Diffuser Cam dataset [27] contains 25,000 paired images captured simultaneously using a standard lensed camera (ground truth) and a mask-based lensless camera Diffuser Cam [1].
Dataset Splits No For the Phlat Cam dataset: 'We use 990 classes for training and 10 classes for testing, following the original protocol.' For the Diffuser Cam dataset: 'These pairs are split into 24,000 images for training and 1,000 for testing.' No explicit mention of a validation dataset split.
Hardware Specification Yes Evrey experiment can be completed within two days using a server equipped with eight NVIDIA RTX 3090 GPUs, a 64-core CPU, 256GB of RAM, and 1TB of storage.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'Stable SR' but does not provide specific version numbers for these or other libraries/frameworks.
Experiment Setup Yes We train the network for 100 epochs with a batch size of 5, using the Adam optimizer [18]. The learning rate is set to 3e-5 for U-Net training in both datasets. For deconvolution kernel training, the learning rate is set at 4e-9 for Phlat Cam and 3e-5 for Diffuser Cam. We set the MSE and LPIPS loss weights to be 1 and 0.05, respectively.