Shadow Generation for Composite Image in Real-World Scenes

Authors: Yan Hong, Li Niu, Jianfu Zhang914-922

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on our DESOBA dataset and real composite images demonstrate the effectiveness of our proposed method.
Researcher Affiliation Academia Yan Hong1, Li Niu1*, Jianfu Zhang2 1 Mo E Key Lab of Artificial Intelligence, Department of Computer Science and Engineering Shanghai Jiao Tong University, Shanghai, China 2 Tensor Learning Team, RIKEN AIP, Tokyo, Japan
Pseudocode No The paper describes the network architecture and procedures in detail using text and diagrams (Figure 3), but does not include explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our dataset and code are available at https://github.com/bcmi/Object-Shadow-Generation Dataset-DESOBA.
Open Datasets Yes First, we contribute a real-world shadow generation dataset DESOBA by generating synthetic composite images based on paired real images and deshadowed images. Our dataset and code are available at https://github.com/bcmi/Object-Shadow-Generation Dataset-DESOBA.
Dataset Splits Yes We follow the training/test split in SOBA dataset (Wang et al. 2020). SOBA has 840 training images with 2, 999 object-shadow pairs and 160 test images with 624 object-shadow pairs. Afterwards, we obtain 615 test image pairs, which are divided into two groups according to whether they have background object-shadow pairs. Specifically, we refer to the test image pairs with Background Object-Shadow (BOS) pairs as BOS test image pairs, and the remaining ones as BOS-free test image pairs. The batch size is 1 and our model is trained for 50 epochs.
Hardware Specification Yes We use Pytorch 1.3.0 to implement our model, which is distributed on RTX 2080 Ti GPU.
Software Dependencies Yes We use Pytorch 1.3.0 to implement our model
Experiment Setup Yes After a few trials, we set λS = λI = 10, λP = 1, and λGD = 0.1 by observing the generated images during training. All images in our used datasets are resized to 256 256 for training and testing. We use adam optimizer with the learning rate initialized as 0.0002 and β set to (0.5, 0.99). The batch size is 1 and our model is trained for 50 epochs.