L-CoDe:Language-Based Colorization Using Color-Object Decoupled Conditions

Authors: Shuchen Weng, Hao Wu, Zheng Chang, Jiajun Tang, Si Li, Boxin Shi2677-2684

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our approach outperforms state-of-the-art methods conditioned on captions.
Researcher Affiliation Academia 1 School of Computer Science, Peking University 2 School of Software and Microelectronics, Peking University 3 School of Artificial Intelligence, Beijing University of Posts and Telecommunications 4 Institute for Artificial Intelligence, Peking University 5 Beijing Academy of Artificial Intelligence 6 Peng Cheng Laboratory
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper states, "To make a fair comparison, we retrain these networks using their publicly available code on the reorganized dataset," referring to Manjunatha et al. (2018) and Xie (2018). However, it does not state that the code for L-Co De is publicly available, nor does it provide a link.
Open Datasets Yes Learning-based image colorization methods could be trained using the Image Net dataset (Russakovsky et al. 2015). However, language-based colorization requires a caption describing the grayscale image, so existing language-based methods choose to use image datasets with captions, such as the COCO-Stuff (Caesar, Uijlings, and Ferrari 2018).
Dataset Splits Yes According to Xie (2018), we keep the images whose captions contain adjectives and have 59K training images and 2.4K validation images left.
Hardware Specification Yes Experiments were conducted on two NVIDIA GTX 1080Ti GPUs.
Software Dependencies No The paper mentions software components like "Bi-LSTM" and "VGG16-BN" but does not specify their version numbers or the versions of any other software dependencies (e.g., Python, PyTorch).
Experiment Setup Yes We set the batch size to 16, α = 0.1 in the ATM. We minimize our objective loss using Adam optimizer with learning rate set as 2 10 4 and momentum parameters β1 = 0.99 and β2 = 0.999.