ThasnimolVSam22d007 erik no i was asked, what is the purpose that link i couldnt find anything there okay images am saving actually, in that color camera code but size is set to 300, when am trying to change it it is showing error only not changing the size? is possible to change the size of image or fixed??i mean is it possible to increase the field of view of image from where i should get the objects pretrained in that network obkjects from that model, if you have any links please share here, am checking it also
jakaskerl Hi ThasnimolVSam22d007 Its a link to a pull request which has the code that aims to do decoding/cropping via script node. Your preview img size is (usually) set to whatever your NN expects. You can instead save video/still/isp to get a larger frame size. Info here. Basically have one stream (preview) go through NN, and the second (video) go directly through xlink/encoder to a file. i'm sorry, what object are you referring to? Thanks, Jaka
ThasnimolVSam22d007 jakaskerl means in that face detection code in github , using one model for object detection, so i want to know what all models are already defined in it, if it is not i want to include those things in the model, whether it is possible to add it in the same model,is it possible do you know whether anyone tried this, or any pretrained models with shoes, goggles, person, face and all already available?? also erik i tried that recording code it is saving some .dat file what is the advantage of those files?/ where i can use that??
ThasnimolVSam22d007 erik i tried this today, but that code with preview video is fully consuming cpu, after running the code consumption is incresed to 30,35 and all, so i am trying to save the results, do you have any ideas for it?? how to take output
jakaskerl Hi ThasnimolVSam22d007 You can add your own classes to an already existing/trained model, by retraining it from a checkpoint along with your new training data. I suggest you look up a few tutorials. I've seen shoes, person, face (haven't yet seen one for goggles), but all as separate models. Your best bet is looking for something like yolo, which usually already has ~ 80 classes. Hope this helps, Jaka
ThasnimolVSam22d007 jakaskerl means, currently am using the code which is already given in the example, so it is using mobilenet ssd, so your saying i need train object and add to the same model or like mobilenet i saw yolo, we can use it same way as yolo in code??
jakaskerl Hi ThasnimolVSam22d007 If your model (be it SSD or yolo) supports retraining with new classes, then yes. Otherwise you will have to freshly train those models to detect your objects. Thanks, Jaka
ThasnimolVSam22d007 @jakaskerl , @erik i have one doubt, any custom trained model will work with oak d camera, now am trying hand gesture detection with yolo8, do you have any idea on the process, actually with this oak d any random model will work or any restriction is there?
erik ThasnimolVSam22d007 yes, as long as operations are supported then you can run the model on OAK. For hand gestures, I would strongly recommend building on top of this project: https://github.com/geaxgx/depthai_hand_tracker
ThasnimolVSam22d007 @erik here in this code which model is using?? do you have any custom model example with yolo - then it will be helpful for me , then i can add more things on it??
erik ThasnimolVSam22d007 it's mediapipe, as written in readme (readme is there for a reason 🙂). Yes, we have custom yolo training, conversion and deployment notebooks here: https://github.com/luxonis/depthai-ml-training/tree/master/colab-notebooks
JanCuhel Hi @ThasnimolVSam22d007, the problem lies with the newest version of the ultralytics. In the tools we use a little bit older version, which doesn't support v8DetectionLoss that the newest version of ultralytics uses. We will deploy the fix within the next release. In the meantime, we have updated the YoloV8 training notebook. You can see the updated notebook here. Basically all you need to do is revert to a bit older version by replacing the first code cell with this: $$ %cd /content/ !git clone https://github.com/ultralytics/ultralytics !cd ultralytics && git reset --hard dce4efce48a05e028e6ec430045431c242e52484 %pip install -qe ultralytics $$ I apologize for the inconvenience. Kind regards, Jan
ThasnimolVSam22d007 JanCuhel with that also i tried but in that colab also it is not working (it was showing - no module named ultralytics.yolo - but i installed all the packages also earlier i tried all the github codes and it was showing no error, inorder to do the custom model deployment and model conversion and all installed alot of libraries - so my oak d folder fully showing an error with - reportMissingImports- but when i am installing those modules it is showing already requirement satisfie
JanCuhel @ThasnimolVSam22d007 I apologize for the delay in my reply, could you please share with me your trained model from before (the one that couldn't be exported)? I will look into exporting the model and also into your issue with the Colab. Best, Jan
ThasnimolVSam22d007 JanCuhel hi @JanCuhel @erik hi i finished the post processing steps up-to making blob file , then i tried to replace it with - https://github.com/thedevyansh/oak-d - this code blob file - my model is yolov8 for face so i changed the label according to that, - but it is throwing error ,and not able to detect the face , do you have any examples with yolo models working with oak d??
erik Hi ThasnimolVSam22d007 , Looks like your model generates too much data, and OAK is unable to parse it. Usually, folks limit outputs to eg. max 100 detections, for which it would consume.. 7 (values per det) * 100 (num of dets) * 2 (FP16->Bytes) => 1.4kB , which is far less than 140kB your model is producing.
ThasnimolVSam22d007 erik 1. our blob file is 52.3 mb and the code is already using 14.5mb blob file - then why it is not running - i didnt get your calculation? is it possible to run yolo converted blob with the mobilenet code? but it is throwing error?(in that one line is there - nn = pipeline.createMobileNetDetectionNetwork() ) , thatswhy i aksed any separate demo code is there to run yolo blob model when i tried with a small dataset also same error is coming<?(6.2mb) here 100 detection u mean ? no of labels or no of images? initially i took (https://app.roboflow.com/iit-madras-3gim0/face-detection-in-all-angles/1) this dataset and tried then with smaller dataset also - but same error is coming in bothcases
erik Also , you are trying to parse YOLO model with MobileSSD node - those are two different nodes. Please see https://github.com/luxonis/depthai-experiments/tree/master/gen2-yolo
ThasnimolVSam22d007 erik this code i tried for yolov8 but that calculation part didn't understand( i trained it with 416 , 416 size)