Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

There is a significant difference in accuracy. #335

Open
chenin-wang opened this issue Dec 4, 2023 · 11 comments
Open

There is a significant difference in accuracy. #335

chenin-wang opened this issue Dec 4, 2023 · 11 comments

Comments

@chenin-wang
Copy link

Despite the bounding boxes (bbox) being visually close in the image, there is a significant difference in their locations.How can this issue be improved?

004001

image

@TontonTremblay
Copy link
Collaborator

What are the cuboid sizes you pass in?

@chenin-wang
Copy link
Author

6x6x6cm
image

@chenin-wang
Copy link
Author

Additionally, I have encountered another issue where the camera can only detect objects within a very narrow distance range. Once the distance becomes too small or too large, detection fails. I have already adjusted the 'location' field values in my dataset across a wide range. I would appreciate any suggestions you could provide. Thank you very much!

@chenin-wang
Copy link
Author

I used isaac sim.

@chenin-wang
Copy link
Author

@TontonTremblay

@TontonTremblay
Copy link
Collaborator

can you put 60,60,60. I am not sure what the scale you used in your dataset. But play with this, you should be able to get to the right location. Also the quaternion might not look the same, isaacsim is probably in opengl coordinate frame, DOPE uses opencv.

@chenin-wang
Copy link
Author

Thank you very much for your response. I will conduct the testing.

@TontonTremblay
Copy link
Collaborator

I will share this thread with the person that did the isaac-sim port.

@chenin-wang
Copy link
Author

chenin-wang commented Dec 11, 2023

Let me clarify my question again.I am facing an issue with object detection in a real-world scenario when adjusting the distance between the camera and the target. Despite resizing the dataset images to 1080x1080 pixels and configuring MIN_DISTANCE as 2.6, MAX_DISTANCE as 3.2, and OBJECT_SCALE as 0.0035, the detection performance does not meet expectations. Previously, you suggested transforming it to 60cmx60cmx60cm. Could you please clarify if you meant resizing the physical dimensions of the object to 60cm or adjusting the OBJECT_SCALE value within the dataset so that the object appears bigger in the image? Additionally, the actual size of my object is only 6cmx6cmx6cm.
this is my Isaac config file for dope, please check it.
dope_config.md

@andrewyguo
Copy link
Contributor

@chenin-wang, could you elaborate more on why OBJECT_SCALE is 0.0035?

What happens if you change this part of the config to:

OBJECT_SCALE:
- 1
- 1
- 1
OBJECT_MASS: 1

@chenin-wang
Copy link
Author

@chenin-wang, could you elaborate more on why OBJECT_SCALE is 0.0035?

What happens if you change this part of the config to:

OBJECT_SCALE:
- 1
- 1
- 1
OBJECT_MASS: 1

the object will be large in the picture

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants