setAnchorMasks missing in v3

TedHerman

Hello, I'm trying to get a Yolo model working with v3, which no longer has the YoloDetectionNetwork node. The docs suggest just using a DetectionNetwork and then a YOLOExtendedParser node. I've had to guess a lot at how some of the parameters work, but one thing has me stumped: when I did a debugging test, I got this error:

[14442C10F178BED200] [1.2] [2.943] [DetectionParser(8)] [error] Mask is not defined for output layer with width '52'. Define at pipeline build time using: 'setAnchorMasks' for 'side52'.

Fine, except that YOLOExtendedParser does not appear to have a setAnchorMasks method (unlike the v2 YoloDetectionNetwork node). I tried finding it with:

import depthai_nodes.node

help(depthai_nodes.node)

There is a setAnchors method, which appears to work, but no setAnchorMasks method. What do I use instead?

KlemenSkrlj

Hi @TedHerman ,
Could you share wich version of YOLO model are you using? If it is one of ones listed here you should be able to use YOLOExtendedparser from depthai-nodes instead of the native DetectionParser.
CC: @MatevzMorato on the backcompatibility for setAnchorMasks on DAIv3 with DetectionParser

TedHerman

KlemenSkrlj Sorry, let me back up to an earlier attempt to understand the problem. I'm using a YOLOv7 model that works with version 2 of the Luxonis stack. The docs say I should, instead of using YoloDetectionNetwork, use DetectionNetwork. So I try that, but see this error (similar to what I initially posted).

[14442C10F178BED200] [1.2] [5.815] [DetectionParser(8)] [error] Mask is not defined for output layer with width '52'. Define at pipeline build time using: 'setAnchorMasks' for 'side52'.

The error message tells me to use setAnchorMasks -- and I have the parameters for anchor masks because I used them with the v2 API and YoloDetectionNetwork. There is no setAnchorMasks method for DetectionNetwork. I'm not sure whether I should try YOLOExtendedParser or not.

I hope that helps clarify the situation.

Update: the AI tells me that I should be putting all the anchor parameters in a JSON file and have that in the same path as the blob. Is this just an hallucination, or is it required in the v3 api?

MatevzMorato

Hey @TedHerman - we haven't binded setAnchorMasks on v3 as it's not mostly expected for people to use NNArchive which has this metadata embedded inside of it.

We'll add bindings to develop, in the mean time you can use:
Camera -> NeuralNetwork -> DetectionParser and then you'll have all the old API exposed directly on the DetectionParser node.

Let me know if this is clear enough, otherwise we can provide a quick example.

TedHerman

MatevzMorato So you are saying that using NNArchive won't work (because sentAnchorMasks is not binded in v3). By the way, I can't find examples of using NNArchive with a local tar.xz in the v3 Python examples. I think not all ways of using v3 with Python are illustrated in examples - probably they are available using C++.

I can instantiate the DetectionParser node though there is no official documentation. I will try that and update this reply with results.

KlemenSkrlj

I can't find examples of using NNArchive with a local tar.xz in the v3 Python examples

You can do this by making the NNArchive object like this and passing it to the NeuralNetwork node. Example is here

nn_archive = dai.NNArchive('./my_nn_archive.tar.xz')
nn = pipeline.create(dai.node.NeuralNetwork).build(cam_out, nn_archive)

And you feed the output of the NN node to the DetectionParser node input

TedHerman

I did not try NNArchive yet. I did try DetectionParser and it has all the bindings. The only difference so far is that I needed to add a call

parsingNetwork.setInputImageSize(416,416)

This avoids the program having many warning messages.

TedHerman

Now I'm trying another model, this one yolov8. The program fails with the message. One other point of difference about this yolov8 model, it's a 640x640 model, if that is relevant.

[2025-09-15 14:40:12.408] [depthai] [error] Device with id 14442C10F178BED200 has crashed. Crash dump logs are stored in: bedrail.v3/.cache/depthai/crashdumps/28fda1d6fc43f5695f24993f2fecda8149462783/crash_dump.json - please report to developers.

[14442C10F178BED200] [1.2] [4.183] [NeuralNetwork(2)] [warning] Network compiled for 4 shaves, maximum available 13, compiling for 6 shaves likely will yield in better performance

[14442C10F178BED200] [1.2] [1757965215.344] [host] [warning] Reconnection successful

[2025-09-15 14:40:17.991] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_3_out' (X_LINK_ERROR)'

[2025-09-15 14:40:17.991] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_0_preview' (X_LINK_ERROR)'

[14442C10F178BED200] [1.2] [1757965219.845] [host] [warning] Closed connection

[14442C10F178BED200] [1.2] [1757965219.845] [host] [warning] Attempting to reconnect. Timeout is 10000ms

TedHerman

One other note on the yolv8 model. According to Ultralytics (this model was trained using their package) Yolov8 does not use anchor masks. Yet when I tried to run it, I got this error message:

[14442C10F178BED200] [1.2] [1.712] [DetectionParser(3)] [error] Mask is not defined for output layer with width '8400'. Define at pipeline build time using: 'setAnchorMasks' for 'side8400'.

[14442C10F178BED200] [1.2] [1.782] [DetectionParser(3)] [error] Mask is not defined for output layer with width '8400'. Define at pipeline build time using: 'setAnchorMasks' for 'side8400'.

I got the same error trying just DetectionNetwork instead of NeuralNetwork -> DetectionParser.

It was after I added the setAnchorMasks (for side8400 and two others) that the crash occurred.

Matija

TedHerman

Hi Ted, how did you compile the model? Easiest way would be to go through HubAI and compile Yolo model using .pt weights directly.

TedHerman

Hey @Matija I think you're right, I should be investigating the model, how it was compiled and work on that - thank you.

Matija

TedHerman

Great, after you compile it through HubAI, it should be straightforward to use it with our generic-example.