After I restore the scene to a specific configuration according to the JSON file, I check the rendered image from the ai2thor and the raw images in the dataset directory. The color and light conditions seem to be different. As I use the raw images to train a policy model, the gap between the dataset and simulators will influence the performance. Therefore, should I design the policy model to deal with this gap or I am missing some details for scene initialization? Here is an example:
And here is the difference of the two images in a single-channel image.
After I restore the scene to a specific configuration according to the JSON file, I check the rendered image from the ai2thor and the raw images in the dataset directory. The color and light conditions seem to be different. As I use the raw images to train a policy model, the gap between the dataset and simulators will influence the performance. Therefore, should I design the policy model to deal with this gap or I am missing some details for scene initialization? Here is an example:
And here is the difference of the two images in a single-channel image.
![image_sub](https://user-images.githubusercontent.com/115710611/221814900-bdb3cdd2-a37e-4e69-9970-4733b7f2e959.jpg)