https://github.com/cvir-lab/hdps
Science Score: 13.0%
This score indicates how likely this project is to be science-related based on various indicators:
-
○CITATION.cff file
-
✓codemeta.json file
Found codemeta.json file -
○.zenodo.json file
-
○DOI references
-
○Academic publication links
-
○Academic email domains
-
○Institutional organization owner
-
○JOSS paper metadata
-
○Scientific vocabulary similarity
Low similarity (9.9%) to scientific vocabulary
Repository
Basic Info
- Host: GitHub
- Owner: CVIR-Lab
- License: mit
- Default Branch: main
- Size: 3.91 KB
Statistics
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
- Releases: 0
Metadata Files
README.md
HDPS
The repository contains the code and dataset in HDPS paradigm, which is currently being submitted to Nature Machine Intelligence. The code is coming soon! we proposed a new paradigm in deep learning, High-Dimensional Projection Selection(HDPS), which is applicable to most network architectures. HDPS integrates the relationship between the target manifold and the input manifold, merging their information interactions to construct a high-dimensional manifold. The foundational step in constructing this high-dimensional manifold lies in identifying the most fundamental properties inherent to the target domain. With that, we then seek to establish a connection between the source manifold and the target manifold, where we finally got a high-dimensional manifold. Through this high-dimensional manifold space, we establish a more effective mapping between the probability distributions of the given target and the input data. Notably, the proposed paradigm introduces several additional nodes to the traditional end-to-end neural network structure, which alleviates two key issues that have long constrained research in computer vision and computational imaging: the challenge of fitting complex distributions with limited samples and the unavoidable information-loosing in initial values.
Performance on NLOS datasets
In the images we present, (a) is the input and (b) is the target. (c) and (d) show the reconstruction results from HDPS, where (c) represents f(z) and (d) represents f(f(z)).
NLOS-MNIST
NLOS-passive
The figure shows the performance on NLOS scene with reconstructing model-face in gray image.
The figure shows the performance on NLOS scene with reconstructing model-face in RGB image.
Performance on Low Light Image Enhancement
Here we present some video-result on sdsd datasets.
sdsd-indoor
sdsd-outdoor
The video on the left depicts the original low-light scene, the middle video shows the enhanced results achieved by HDPS, and the rightmost video serves as the ground truth.
We reorganized the dataset page, please see the new official dataset NLOS of HDPS for detail.
Owner
- Name: CVIR-Lab
- Login: CVIR-Lab
- Kind: organization
- Repositories: 2
- Profile: https://github.com/CVIR-Lab
GitHub Events
Total
- Member event: 1
- Push event: 31
- Create event: 2
Last Year
- Member event: 1
- Push event: 31
- Create event: 2