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; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Projected Robust PCA with Application to Smooth Image Recovery
Authors: Long Feng, Junhui Wang
JMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Furthermore, a comprehensive simulation study along with a real image data analysis further demonstrate the superior performance of PRPCA in terms of both recovery accuracy and computational advantage. |
| Researcher Affiliation | Academia | Long Feng EMAIL Department of Statistics & Actuarial Science The University of Hong Kong Pokfulam, Hong Kong; Junhui Wang EMAIL Department of Statistics The Chinese University of Hong Kong Shatin, Hong Kong |
| Pseudocode | Yes | Algorithm 1: Proximal gradient for PRPCA; Algorithm 2: Accelerated proximal gradient for PRPCA |
| Open Source Code | No | The paper does not contain any explicit statement about open-sourcing the code, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We consider the gradyscale Lenna image, which is of dimension 512x512 and can be found at https://www.ece.rice.edu/~wakin/images/. In our study, we consider the scenario with one person walking on straight line. The video is half-resolution PAL standard (288*384 pixels, 25 frames per second), with a total of 611 frames. From computational perspectives, we first resize each frame of video into 56 x 80 (4480) pixels and then keep the last 411 frames. We stack each frame into a matrix column and obtain a large matrix of dimension 4480x 411. Denote the large matrix as Theta. As in the Lenna image analysis, we consider recovering Theta from noisy observations Z = Theta + E, with E ~ N(0, sigma^2) and the same set of sigma: sigma = 0.05, 0.1, 0.15, 0.2, 0.25. Such setting refers to the case that each frame is compromised with independent white noise. |
| Dataset Splits | No | For the simulation study, the data (low-rank matrix X0 and sparse component Y0) were generated synthetically, not split from an existing dataset. For the real data analysis, the paper performs image recovery on the full image/video, rather than splitting it into training/test/validation sets for model training/evaluation. |
| Hardware Specification | Yes | Finally, we report the required computation time (in seconds; all calculations were performed on a 2018 Mac Book Pro laptop with 2.3 GHz Quad-Core Processor and 16GB Memory). |
| Software Dependencies | No | The paper mentions using a 'proximal gradient algorithm' and its 'accelerated version' (Nesterov, 2013), and references specific techniques like 'Singular Value Thresholding' (Cai et al., 2010) and 'Soft Thresholding operators', but does not specify any particular software libraries, packages, or their version numbers that were used for implementation (e.g., Python, PyTorch, TensorFlow, MATLAB with specific toolboxes/versions). |
| Experiment Setup | Yes | The other parameters are fixed at r = 10 and sigma0 = 0.6 if not otherwise specified. For all four sets of (P, Q), we use the same penalty level with lambda1 = sqrt(2N)sigma and lambda1 = sqrt(2)sigma. The simulation is run over a grid of values for the matrix dimension N, noise level sigma and sparsity level rhos. sigma = 0.2, 0.4, 0.6, 0.8, 1; rhos = 0.05, 0.1, 0.15, 0.2, 0.25; N = 60, 100, 200, 300, 400. |