Datasets:

pandaphd

/

camera_settings

like 9

Hi Yuxin, thank you for your attention. Here, we only provide preprocessed camera setting gallery images, which serve as the base for generating subsequent contrastive data. The camera metadata is not always necessary for this project. The approach adopted in this work relies more on the physical simulations of different settings (detailed guidance can be found in the supplementary materials of our paper: https://arxiv.org/abs/2412.02168. We will release the code for the physical simulations at a later stage.

Hey, @pandaphd I am also curious about the dataset and the physics-based approach you took to simulate the camera parameters and color temperatures. If i understand your supplementary information correctly, you were applying a bunch of physics rules to build contrastive data from each base image, and then leverage these contrastive data to train a contrastive camera encoder to understand these rules.

Have you ever compared your approach with a much more naive approach: generating an image with base model, and then apply those physics rules to transform them into new images with different camera/color temperature settings?

You hit the points.

For certain tasks (e.g., color temperature), post-processing is feasible. However, for other tasks such as focal length, it requires center cropping followed by super-resolution on a generated result that may have incorrect perspective relationships. In such cases, post-processing becomes more cumbersome and uncontrollable. Our approach, on the other hand, unifies multiple tasks in an end-to-end manner, minimizing the creation burden to the greatest extent.