TensorFlow example - resolve incorrect flower labelling about depthai-experiments HOT 13 CLOSED

I found the issue thanks to Szabi from Discord.

Apparently, we have to add these two lines or the color information won't match with the OpenVINO format.

cam_rgb.setBoardSocket(dai.CameraBoardSocket.RGB)
cam_rgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.RGB)

Then, I have to reverse the color again

frame_rgb = cv2.cvtColor(frame,cv2.COLOR_BGR2RGB)
# if the frame is available, draw bounding boxes on it and show the frame
if result is not None:
   cv2.putText(frame_rgb, "{} ({}%)".format(result["name"], result["conf"]), (20, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0))

https://colab.research.google.com/drive/1oNxfvx5jOfcmk1Nx0qavjLN8KtWcLRn6?usp=sharing

Also, I have made a validation script here and so you can see the IR format is working correctly. I think that is a very important step as we have to validate our input format (this is a deadly trap as nothing will work unless these are all pair-matched perfectly).

Please re-run this COLAB and download the new blob. And then perhaps make these updates and see if it works. Once all good we can make a new pull request and merge it.

from depthai-experiments.

To test it, we should generate a .mp4 based on the training set? That way we know the ground truth?

I believe we can make a video (like a slide show) with pictures using FFmpeg... =) Just one idea to validate our work =)

-ray

from depthai-experiments.

I was thinking of adding a capability to send the images one by one and present them to the users - will also showcase how to use images instead of videos to interact with DepthAI

from depthai-experiments.

And then, as you suggested @raymondlo84, feed there a training data

from depthai-experiments.

=) Any suggestions on the blob converting step? Would it be related to I'm using 2021.2? And what are the easiest way to validate the blob?

from depthai-experiments.

Great question. @realWadim could you investigate what is going wrong here?

from depthai-experiments.

=) Any updates? I think we will turn this into a tutorial once it is working. Also, would love to try classifying something different for fun too.

from depthai-experiments.

Sorry about the delay. Actually I had forgotten I assigned this to Wadim, who is no longer with Luxonis. I'm going to see if Erik can help debug this.

Thanks,
Brandon

from depthai-experiments.

=) Erik? Ok, would be great to know any sort of feedback. I am gonna ping a few and see if someone get enough insight into that. I am still guessing it has to do with the myriad compile step, and I don't have enough technical understanding of the entire pipeline there.

from depthai-experiments.

Awesome, thank you! And sorry about the delay on this! And thanks for tracking it down!

from depthai-experiments.

The inference works ok now (although dandelion is only recognized in the mature form, but it's still better than before),
PR with related changes is here

from depthai-experiments.

Just tested it today and seems to be working as expected. I think this is sufficient to be a 'demo' now. It has all the puzzle together. =)

-ray

from depthai-experiments.

Thanks @raymondlo84 and also thanks for continuing to push on this and get it resolved! And then also testing/confirming! Super appreciated!

from depthai-experiments.