Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
DSI2I: Dense Style for Unpaired Exemplar-based Image-to- Image Translation
Authors: Baran Ozaydin, Tong Zhang, Sabine Susstrunk, Mathieu Salzmann
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of our method on four datasets using standard metrics together with a localized style metric we propose, which measures style similarity in a class-wise manner. Our results show that the translations produced by our approach are more diverse, preserve the source content better, and are closer to the exemplars when compared to the state-of-the-art methods. Our experiments show both qualitatively and quantitatively the benefits of our method over global, imagelevel style representations. |
| Researcher Affiliation | Academia | Baran Ozaydin EMAIL School of Computer and Communication Sciences, EPFL, Switzerland; Tong Zhang EMAIL School of Computer and Communication Sciences, EPFL, Switzerland; Sabine Süsstrunk EMAIL School of Computer and Communication Sciences, EPFL, Switzerland |
| Pseudocode | No | The paper describes the proposed method and mathematical formulations, such as the Sinkhorn algorithm, but it does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Project page: https://github.com/IVRL/dsi2i |
| Open Datasets | Yes | We evaluate our method on real-to-synthetic and synthetic-to-real translations using the GTA Richter et al. (2016), Cityscapes Cordts et al. (2016), and KITTI Geiger et al. (2012) datasets. We also evaluate our method on real-to-real translation using the sunny and night splits of the INIT Shen et al. (2019) dataset. [...] we provide results on the summer2winter and monet2photo datasets Zhu et al. (2017), in both directions. |
| Dataset Splits | No | The paper mentions using well-known datasets like GTA, Cityscapes, and KITTI, and refers to existing setups ('We use the same setup as previous works'). While these datasets often have standard splits, the paper does not explicitly provide the specific training, validation, or test splits (e.g., percentages or counts) for these datasets within its text. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types) used for running its experiments. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers, such as programming language versions or specific library versions (e.g., PyTorch 1.x, TensorFlow 2.x), used in their implementation. |
| Experiment Setup | Yes | Images are resized to have a short side of 256. We borrow the hyperparameters from Huang et al. (2018) but we scale the adversarial losses by half since our method receives gradients from three adversarial losses for one source image. We do not change the hyperparameters for the perceptual losses. The entropy regularization term in Sinkhorn s algorithm in Eq. 12 is set to 0.05. During training, we crop the center 224x224 pixels of the images. All the models are trained for 800K iterations with 224x224 images. We use linear learning rate decay after 400K iterations as suggested in these works. |