Task-Specific Scene Structure Representations

Authors: Jisu Shin, Seunghyun Shin, Hae-Gon Jeon

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct a variety of experiments on low-level vision tasks, including self-supervised joint depth upsampling (Sec.4.1) and unsupervised single image denoising (Sec.4.2), to demonstrate the effectiveness of our SSGNet.
Researcher Affiliation Academia Jisu Shin*, Seunghyun Shin*and Hae-Gon Jeon AI Graduate School, GIST, South Korea {jsshin98, seunghyuns98}@gm.gist.ac.kr, haegonj@gist.ac.kr
Pseudocode No The paper describes the network architecture (Fig. 2) and loss functions, but it does not provide any pseudocode or algorithm blocks.
Open Source Code Yes Our source codes are available at https://github.com/jsshin98/SSGNet.
Open Datasets Yes Prior to the evaluations, we train our SSGNet on a well-known NYUv2 dataset (Silberman and Fergus 2011), consisting of 1,000 training images and 449 test images.
Dataset Splits No The paper mentions training and test sets for NYUv2 but does not specify a separate validation set or describe how validation was performed for any dataset used.
Hardware Specification Yes The training on SSGNet took about 40 hours on two NVIDIA Tesla v100 GPUs.
Software Dependencies No The paper mentions 'public Pytorch' but does not specify a version number or any other software dependencies with their versions.
Experiment Setup Yes The learning rate and the batch size are set to 0.0001 and 4 on SSGNet, respectively. We train the proposed framework on images with a 256 256 resolution. ... the hyperparameter γ is set to 0.9 in our implementation. where λ is the hyper-parameter, and is empirically set to 40.