Models does not detect objects CHIP_ID=12 SDK=1.2.1 about acap-computer-vision-sdk-examples HOT 11 CLOSED

The model is spaghettinet.
Could it be an unsupported operation or something like that?

from acap-computer-vision-sdk-examples.

Hi @sheh
We would have to investigate this issue, and it might take time.
Could you share with us a demo version of your model already trained (non with sensible data)?

from acap-computer-vision-sdk-examples.

Model file.

I slightly modified object-detection-python example to print scores.
I built the docker images then uploaded to the device and run:

./docker-compose-acap --env-file ./config/env.$ARCH.artpec8 up

get no detection:

object-detector-python_1  | Time for call to inference-server: 136 ms
object-detector-python_1  | 0 Objects found
object-detector-python_1  | Total time for image inference: 154 ms
object-detector-python_1  | Time for call to inference-server: 50 ms
object-detector-python_1  | 0 Objects found
object-detector-python_1  | Total time for image inference: 62 ms
object-detector-python_1  | Time for call to inference-server: 52 ms
object-detector-python_1  | 0 Objects found
object-detector-python_1  | Total time for image inference: 60 ms
object-detector-python_1  | Time for call to inference-server: 51 ms
object-detector-python_1  | 0 Objects found

then I just change artpec8 to cpu and get detection:

./docker-compose-acap --env-file ./config/env.$ARCH.cpu up

object-detector-python_1  | Total time for image inference: 256 ms
object-detector-python_1  | Time for call to inference-server: 237 ms
object-detector-python_1  | 3 Objects found
object-detector-python_1  | dog 0.809
object-detector-python_1  | bicycle 0.793
object-detector-python_1  | truck 0.680

docker-compose-acap:

#!/bin/bash

set -x
set -e

TSL_ROOT=...

docker-compose --tlsverify --tlscacert=$TSL_ROOT/ca.pem \
        --tlscert=$TSL_ROOT/cert.pem \
        --tlskey=$TSL_ROOT/key.pem \
        -H tcp://${AXIS_HOST}:2376 $@

Also I added IMAGE_PATH=dog416.png to docker-compose.yml (static-image.yml looks outdated).

from acap-computer-vision-sdk-examples.

Could you try to upgrade your firmware to 10.12 and see if the problem persist?

from acap-computer-vision-sdk-examples.

I did that - the same result.

from acap-computer-vision-sdk-examples.

Any thoughts about cause of the issues and what should I check? It blocks me from release an app.

from acap-computer-vision-sdk-examples.

Hi @sheh

I've been in touch with the team that will work on this issue to have an idea of when this issue will be handled.
Unfortunately, it is a complex problem, and we don't have the resources to check your specific model soon, it will probably be pushed until later than one month from now.

Probably the easiest solution for you is to stay with the firmware 10.8 until this gets fixed.

from acap-computer-vision-sdk-examples.

Thanks! Got it.
Just a guess - could it be because of relu6 activation? There are not such activation in the model example but I have them in my model.

from acap-computer-vision-sdk-examples.

It might be, but not sure, also I don't see the relu6 on netron.
We have a similar problem with the pose estimation example that uses Movenet.
using the accelerator after FW 10.8 all the confidences printed are zeros.

from acap-computer-vision-sdk-examples.

HI @sheh
We recently released a new firmware version for our camera (11.0).
It solved some similar issues with other models, where the inference was being executed, but the output was corrupt.
I suggest you to try your model with the latest firmware, to see if it solves also your issue.

from acap-computer-vision-sdk-examples.

Hello again @sheh

I confirm that with the new firmware the problems seems to be fixed
I can detect the dog, car and truck in the image running it on the accelerator.

Feel free to open a new issue or discussion if you have further questions

from acap-computer-vision-sdk-examples.