Soft Superpixel Neighborhood Attention

Authors: Kent W Gauen, Stanley H. Chan

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

Reproducibility Variable Result LLM Response
Research Type Experimental SNA outperforms alternative local attention modules on image denoising, and we compare the superpixels learned from denoising with those learned with superpixel supervision. This section demonstrates the impressive benefit of superpixel neighborhood attention compared to standard neighborhood attention. To verify whether the improvement is due to the proposed method, we compare several variations of both attention methods. Section 5.2 compares different attention modules within a simple network architecture on Gaussian denoising, which empirically verifies the theoretical findings in Sections 4.3 and A.2. Section 5.3 compares the superpixel probabilities learned from the denoising loss function with superpixels learned through supervised training.
Researcher Affiliation Academia Kent Gauen Purdue University gauenk@purdue.edu Stanley Chan Purdue University stanchan@purdue.edu
Pseudocode No The paper contains equations like (3), (4), (5) for NA, H-SNA, and SNA, and describes the steps, but not in a pseudocode format.
Open Source Code Yes 1Code for this project is available at https://github.com/gauenk/spix_paper
Open Datasets Yes We train each network for 800 epochs using a batch size of 2 on the BSD500 dataset [36] using a learning rate of 2 10 4 with a decay factor of 1/2 at epochs 300 and 600. Testing datasets are Set5 [44], BSD100 [36], Urban100 [45], and Manga109 [46].
Dataset Splits No We train each network for 800 epochs using a batch size of 2 on the BSD500 dataset [36]. Testing datasets are Set5 [44], BSD100 [36], Urban100 [45], and Manga109 [46].
Hardware Specification Yes The code is implemented in Python using Pytorch, Numpy, Pandas, and CUDA and run using two NVIDIA Titan RTX GPUs and one RTX 3090 Ti GPU [40 43].
Software Dependencies Yes The code is implemented in Python using Pytorch, Numpy, Pandas, and CUDA and run using two NVIDIA Titan RTX GPUs and one RTX 3090 Ti GPU [40 43]. Cuda, release: 10.2.89, 2020.
Experiment Setup Yes We train each network for 800 epochs using a batch size of 2 on the BSD500 dataset [36] using a learning rate of 2 10 4 with a decay factor of 1/2 at epochs 300 and 600. The network is optimized with Adam [39].