I can see the skeleton of an adult, but if I want to see the baby skeleton, even if I modfiied these two lines
obj_runtime_param.detection_confidence_threshold = 0.01
, the baby still can not be recognized easily. What approach shall I take in order to improve the baby skeleton recognition?
Or I need to find baby skeleton model and try to merge with your code? (Sounds complicated…)
Our AI model has been mostly trained with adults. The detection of babies might be less accurate.
I’ll suggest you to disable the “body fitting” option, it might help a bit for baby detection.
I just tried to comment out the body fitting, I don’t see the effect though…
The baby skeleton keeps flicking, but the cam is just looking at the same scene without anything moving. (Baby doll) object_confidence is around 9% sometimes 70%
the object_id increment fast, is there any parameter to recognize the same object in the same location?
Note1: obj_runtime_param.detection_confidence_threshold is 0.01#40 (very low already, is there a threshold for object and another threshold for skeleton?)
Note2: I have kind of confuse with other threshold like runtime threshold, does it affect the cloud not object?
In my application, the baby does not move very often, is there any setting that once the skeleton is detected, it shows up in the 2D window for longer time? And without changing the object_id?
Plan B, is there any tutorial link of how to train a baby skeleton? And how to import the model into your code?
Do you need to detection the skeleton of the baby or only it’s position (bounding box) ?
Because, with the SDK you can use the Custom Detector feature (Using the Object Detection API with a Custom Detector | Stereolabs) which allow you to ingest 2D detection data from an external object detector into the SDK that will compute the 3D data (such as position, 3D bounding box…). That allow you to use another AI model, that can potentially detect baby more accurately.
Plan A is to detect the baby skeleton. Is it hard? I can learn.
As what I understand from your message,
Plan B is use Yolo - custom detect for the baby? I train a yolo model for it?
It depends what kind of information you need.
If the bounding box is enough for your project, you can either try our object detection model (called MULTI_CLASS_BOX_ACCURATE) or use an external detector such as YOLO.
We will do more tests internally for the baby skeleton detection and try to improve the quality of the detection in the future release.
Thx for the information.
Pls let me know if there is any new release.
Here is an example.
Many times, the baby cannot be recognised.
Even if the baby is seen, the skeleton is wrong.
Could you record one or multiple SVOs showing this issue and share them with us.
Send us a mail at firstname.lastname@example.org with the SVOs files.
It will help us a lot.
Pls check email.