The observation images regenerated by the dataset_state_to_obs script are upside down.

[pumfish]

Hi，I'm using the Robomimic to regenerate demonstration by dataset_state_to_obs.py, follow this https://robomimic.github.io/docs/datasets/robosuite.html#extracting-observations-from-mujoco-states
But I found that the regenerate image, eg., frontview_image is upside down.
I have set ```macros.IMAGE_CONVENTION="opencv"````, but it's uesless. Is there something I haven't set up correctly?
Below is my scripts

 python dataset_states_to_obs.py --dataset ${DATA_DIR} \
         --output_name "add_agentview_demo.hdf5" \
         --done_mode 0 --shaped \
         --camera_names frontview agentview --depth \
         --camera_height 256 --camera_width 256 \

By the way, I also found that the depth vaule is different from the original. for example, the original depth from frontview_depth is from 0.992 to 0.997, but the regenerate depth value is from 1.3 to 3.8

How can I solve these issues?

[amandlek]

What version and branch of robosuite and robomimic are you on?

As for the depth value, we convert the normalized values into real-valued depth that corresponds to distances here, since this is the desired behavior for most applications. You can toggle that behavior if you'd like to go with the normalized values.

[pumfish]

Thanks for your reply
The version of robosuite is 1.4.1, robomimic is 0.3.0
When I generate point clouds using inverted images and normal images, I observe misalignment even though the depth values have been converted to real-values. I would like to know if the inverted images are the cause of the misalignment in the point clouds.
The code for synthesizing point clouds is referenced https://github.com/haonan16/Stow/blob/b3d3045a64992a190bbad9f25b14e83b45d2ae8b/perception/sample.py#L46-L120