OoDHDR-Codec: Out-of-Distribution Generalization for HDR Image Compression

Authors: Linfeng Cao, Aofan Jiang, Wei Li, Huaying Wu, Nanyang Ye158-166

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results show that the Oo DHDR codec achieves strong competitive in-distribution performance and state-of-the-art Oo D performance.
Researcher Affiliation Collaboration Linfeng Cao1, Aofan Jiang1, Wei Li2, Huaying Wu1, Nanyang Ye1* 1 Shanghai Jiao Tong University 2 Huawei Noah s Ark Lab
Pseudocode No The paper does not contain any explicit pseudocode blocks or algorithms labeled as such.
Open Source Code Yes Code available at https://github.com/caolinfeng/Oo DHDR-codec.
Open Datasets Yes The environment esdr is constructed by SDR images from DIV2K (Agustsson and Timofte 2017) and Flickr2K (Lim et al. 2017) datasets with around 3500 samples. To construct ehdr, we collect HDR images from (Funt and Shi 2010; Kalantari and Ramamoorthi 2017; Yeganeh and Wang 2013; Narwaria et al. 2013; Debevec and Malik 2008; Ward et al. 2006), pfstools resources 2, HDRI HEVEN, with 480 samples in total and all of them are in Radiance RGBE (.hdr) (Ward 1991) format that contains the absolute luminance value (cd/m2).
Dataset Splits No The paper mentions training on DIV2K and Flickr2K datasets and testing on Kodak and HDRI HAVEN, but does not explicitly specify a separate validation dataset split or its size/usage.
Hardware Specification Yes Our framework is implemented with Py Torch 1.6.0, CUDA v11.4 on NVIDIA 2080Ti GPU.
Software Dependencies Yes Our framework is implemented with Py Torch 1.6.0, CUDA v11.4 on NVIDIA 2080Ti GPU.
Experiment Setup Yes The model is trained up to 250 epochs with initial learning rate of 1e 3, which decreases to 1e 4 at 200th epoch. To obtain the models with different rate control, the hyper-parameter λ1 is chosen from [12, 40, 150, 300], and the channel number of latent representation is fixed at 192. For each model, the hyperparameter of λ2, λ3 and λ4 is set to 0.95, 1e 5 and 1, the segment number of TMO is set to 10.