yolov7 custom tiny model: X_LINK_ERROR | side values? | poor detection with OAK-D

[ZucchiniAI]

Hello,
I have converted a custom trained yolov7 tiny model (13 classes and mAP=0.75, 1024x1024) into a blob with http://tools.luxonis.com/
I have used the blob following the instruction of YoloV7_training.ipynb notebook from [depthai-ml-training] repo with a OAK-D camera (connected to USB3 with/or without additional power supply)
I have 2 issues:

1) after few seconds I get an error. Why?
Traceback (most recent call last):
File "main.py", line 51, in
pv.prepareFrames()
File "/home/mz/Projects/ObjectDetection/depthai/depthai_sdk/src/depthai_sdk/managers/preview_manager.py", line 148, in prepareFrames
packet = queue.tryGet()
RuntimeError: Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: 'color' (X_LINK_ERROR)'

2) before the error I see very few and bad detections even if the trained model gave very good results on the test set for static images (mAP~0.75). Why? Please see my Obs below: is that the reason?

Obs: The JSON note in https://github.com/luxonis/depthai-experiments/tree/master/gen2-yolo/device-decoding is not clear to me:
I have not changed the Json file of my model which is 1024x1024 as I do not understand WHERE I have to change the "side" entries: I have no side32 or side16 but many of them and all concerning the anchor masks. See attached my json file from the blob conversion: shall I do change something? how and where exactly? best.zip

Note: Values must match the values set in the CFG during training. If you use a different input width, you should also change side32 to sideX and side16 to sideY, where X = width/16 and Y = width/32. If you are using a non-tiny model, those values are width/8, width/16, and width/32.

Thank you in advance
Marco

[Erol444]

Hi @ZucchiniAI ,
We apologize for the delay, I'll check this issue tomorrow morning.
Thanks, Erik

[Erol444]

Hi @ZucchiniAI , are you using the latest version of depthai? We have tried the same process, and it crashes with old (2.16) depthai, but not with the newest depthai (2.17.4).
Thanks, Erik

[ZucchiniAI]

Hello Erik,

1. the X_LINK_ERROR (Point1) happened with the depthai 2.17.0.0.
   I have now updated to 2.17.4 and indeed this do not happen at all: thank you!

What about my Point2? I see now that the objects are detected with good confidence but with much redundant bounding boxes (e.g. a couple for each object with different sizes, almost in the region of the ground truth) --> it seems to me the NMS does not work well. Could you please tell if the reason is my supposition (see Obs. in my question above) and how/where exactly shall I change the indicated sideX and sideY for my tiny model (1024x1024)?

Thank you in advance! Marco

[Erol444]

Hi @ZucchiniAI , could you try with different IoU threshold setting?

[ZucchiniAI]

Hello Erik, I was trying as you suggested, but unfortunately the result do not convince me totally :-(

a) a couple of time I got still the X_LINK_ERROR even if env with depthai 2.17.4 was activated .... what is the cause? (I have the USB-C to USB 3 port and the additional power connected): RuntimeError: Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: 'rgb' (X_LINK_ERROR)'

b) By changing the iou_threshold in the best.yaml in range 0.2 - 0.8 I saw some changes (0.2 being better, but not perfect) but anyhow the detections appear and disappear rapidly and in different positions: sometime is very clear that the object is detected twice even if the model before conversion detected one only. Sometime is difficult to say if is a double bounding box for the same object or intermittent small and large: is there a way to reduce the inference frequency (with my tiny yolo7 I get some 11 fps):
My set up is to point the OAK-D camera towards the pc screen where I have the same images which yolov7 before the conversion detected correctly.

c) sorry if I ask but again your hint in https://github.com/luxonis/depthai-experiments/tree/master/gen2-yolo/device-decoding is not clear to me: I have not changed the Json file of my model which is 1024x1024 as I do not understand WHERE I have to change the "side" entries: I have no side32 or side16 but many of them and all concerning the anchor masks. See above attached my json file from the blob conversion: shall I do change something? how and where exactly?

Thank you!
Marco

[Erol444]

Hi @ZucchiniAI ,
a) That's a very generic error, basically means "something went wrong". Is it sporadic error after X hours?
b) Could you share a video? That would help a lot with debugging.
c) Anchor masks are are already written in JSON (that you downloaded from the tools.luxonis.com) and will be loaded to the device, I am not sure why you would need to change them?

Thanks, Erik

[ZucchiniAI]

Hi Erik,
sorry for the delay:

a) well it happens very often and just after minutes ... (I use a Ubuntu 20 and your original cables): Any way to have more precise log?
b) I would do this, but I do not know how: I mean to save the video from OAK with the live-detected objects? Shall I use the managers DepthAI SDK for it? have you an example?
c) Sorry I did want to change really but I thougth I had to as my resolution is different from the standard one: By reading again I understand now that if I keep the same resolution after training I do not need to change anything.

Thanks!
Marco