Giters

geaxgx / depthai_hand_tracker

Running Google Mediapipe Hand Tracking models on Luxonis DepthAI hardware (OAK-D-lite, OAK-D, OAK-1,...)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Using laptop integrated camera

emoullet opened this issue · comments

commented

Hello,

I'm building an application for which I'd like to be able to switch between an OAK-D S2 and a standard, laptop integrated webcam.
It is my understanding that it should possible (at the cost of several unavailable features) by running :

python demo.py -i '0'

But i get the following RunTime error :

Palm detection blob : /home/emoullet/GitHub/depthai_hand_tracker/models/palm_detection_sh4.blob
Landmark blob : /home/emoullet/GitHub/depthai_hand_tracker/models/hand_landmark_lite_sh4.blob
Traceback (most recent call last):
File "/home/emoullet/GitHub/depthai_hand_tracker/demo.py", line 59, in <module>
tracker = HandTracker(
File "/home/emoullet/GitHub/depthai_hand_tracker/HandTracker.py", line 130, in __init__
self.device = dai.Device()
RuntimeError: No available devices

Am I missing something on the arguments to pass for it to work with integrated camera ? Or is there some preparatory work to do on the cam (i.e. calibration) ?

Many thanks in advance, and hats off for this amasing tool !

commented

Hi,
Even if you want to process the video frames coming from your webcam, you still need the Myriad processor of your OAK-D to run the detection and landmark regression models. So you still need the OAK-D be plugged in. The RuntimeError: No available devices is the typical message you get when no OAK device is seen by the system.

commented

Thank you for your answer !
Is there any way to run the detection and landmark regression model directly on the computer ? I already use mediapipe but many of the additionnal features proposed in this repo are quite interesting, and if possible I would rather directly use them rather than re-implement them..

commented

Using directly mediapipe is the best method to run the models on your computer.
Another method is to convert the models to Openvino format and then run them on your Intel CPU. That is actually what I did in this older repo (https://github.com/geaxgx/openvino_hand_tracker) as a preliminary step when creating the repo dedicated to the OAK devices. So initially, https://github.com/geaxgx/depthai_hand_tracker was just a copy of https://github.com/geaxgx/openvino_hand_tracker where I replaced the code specific to Openvino by equivalent code specific to depthai. But because I was focused on OAKD, further modifications and new features were only done on this repo and the openvino repo was kind of abandoned.

commented

can you suggest a way to improve media pipe accuracy to detect hand from long distance in laptop using a method you used for OAK-D in your implementation , Thank you

commented

@udayzee05 If you are working with mediapipe, you can directly use Holistic (https://google.github.io/mediapipe/solutions/holistic). It relies on blazepose to estimate the body pose. And the body pose gives the region where to look for the hands. I am using movenet instead of blazepose but the principle is similar.