Single Image Reflection Separation via Dual-Stream Interactive Transformers
Authors: Qiming Hu, Hainuo Wang, Xiaojie Guo
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results reveal the merits of the proposed DSIT over other state-of-the-art alternatives. |
| Researcher Affiliation | Academia | Qiming Hu, Hainuo Wang, Xiaojie Guo College of Intelligence and Computing, Tianjin University, Tianjin 300350, China |
| Pseudocode | Yes | A.1 Algorithm for Dual-Attention Interactive Block We offer the overall procedure of the DAIB in Alg. 1, corresponding to the description in Sec. 3.1. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/mingcv/DSIT. |
| Open Datasets | Yes | Our training datasets include both synthetic and real-world images. Following [23], we design two data settings for fair comparison: I. 7,643 synthesized pairs randomly sampled from the PASCAL VOC dataset [13] in each epoch and 90 real pairs from [60]. II. 200 extra real pairs from the Nature dataset [29], and 13,700 synthesized pairs sampled from [60] instead. |
| Dataset Splits | Yes | Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: In the Experiments section, we provide detailed descriptions of all the training and testing details required. |
| Hardware Specification | Yes | The learning rate is fixed as 10 4 with a batch size of 1 on a single RTX 3090 GPU. |
| Software Dependencies | No | Our models are all implemented via the Py Torch framework and optimized with Adam optimizer for 20 or 80 epochs based on different data settings. |
| Experiment Setup | Yes | The training image size is fixed as 384 384. The window size of attention mechanisms, NW , is fixed to 12 12, and the number of windows, NT , varies depending on the spatial scale of the features. Our models are all implemented via the Py Torch framework and optimized with Adam optimizer for 20 or 80 epochs based on different data settings. The learning rate is fixed as 10 4 with a batch size of 1 on a single RTX 3090 GPU. |