ILSGAN: Independent Layer Synthesis for Unsupervised Foreground-Background Segmentation
Authors: Qiran Zou, Yu Yang, Wing Yin Cheung, Chang Liu, Xiangyang Ji
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform experiments on various datasets including Cars (Krause et al. 2013), CUB (Wah et al. 2011), and Dogs (Khosla et al. 2011). |
| Researcher Affiliation | Academia | Tsinghua University, BNRist qiranzou@gmail.com, yang-yu16@foxmail.com, zhangyx20@mails.tsinghua.edu.cn, {liuchang2022, xyji}@tsinghua.edu.cn |
| Pseudocode | No | The paper describes the algorithm steps in mathematical formulations and prose, but does not include a dedicated pseudocode or algorithm block. |
| Open Source Code | No | The paper does not explicitly state that the source code for the described methodology is publicly available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate our method on Stanford Cars (Cars) (Krause et al. 2013), Stanford Dogs (Dogs) (Khosla et al. 2011), and Caltech-UCSD Birds 200-2011 (CUB) (Wah et al. 2011). |
| Dataset Splits | No | We follow the train-test split from (Yu et al. 2021) to split Cars and Dogs dataset, leading to 6218 training images and 6104 testing images for Cars as well as 3286 training and 1738 testing images for Dogs. The CUB dataset is split into 10k images for training and 1k images for testing following (Chen, Arti eres, and Denoyer 2019). |
| Hardware Specification | Yes | Experiments are conducted on single RTX 2080Ti GPU. |
| Software Dependencies | No | The paper mentions using Style GAN2 and UNet but does not specify versions for these or other general software dependencies like Python or PyTorch. |
| Experiment Setup | Yes | Hyperparameters are consistently set across datasets: non-empty mask loss weight λm = 2, mask size threshold η = 0.25, mask binarization loss weight λb = 2, and ILSMI loss weight λils = 1. Our ILSGAN is optimized using Adam optimizer with initial learning rate 0.0025 and beta parameters 0 and 0.99 until 8M images have been shown to the discriminator. |