This year has been a difficult one for me in terms of having time to do projects and try out new hardware. Back in mid-February I pre-ordered an Arducam PiNSIGHT which was selling for a $30 discount from its normal $99.99 price. I've had it since the end of May but just haven't had the opportunity to try it out.
The PiNSIGHT is an AI camera module that is designed to piggyback onto an RPi 5/4B/3B+. The AI comes from a Luxonis OAK-SoM which uses an Intel Myriad X VPU. Basically what's in an Intel NCS2 so it will do 4 TOPS. I wasn't aware of the RPi AI Kit when I ordered this.
The camera uses an IMX378 12MP sensor with autofocus so it's like an earlier version of the IMX708 on the RPi Camera Module 3.
The front aluminum shell doubles as a heatsink for the SoM and the PCB assembly attaches to it. The connection to the Pi is via USB jumper cable for data and power.
Hardware Assembly
I added an RPi5 4GB with the active cooling module and for now just printed a backplate so that I have an open frame camera shown below from the front and above.
Software Installation
There is a package installation script that will install all of the required dependencies and the Luxonis DepthAI example directories. Of course, this is only a single camera so the depth examples will not work but all the detection ones will.
wget https://github.com/ArduCAM/arducam_ppa/releases/download/v1.0.2/pinsight_install_dependencies.sh chmod +x pinsight_install_dependencies.sh ./pinsight_install_dependencies.sh
Demo Applications
There is a demo application depthai_demo.py that will run object detection by default but can run additional models using the -cnn flag.
Available models:
- deeplabv3p_person
- face-detection-adas-0001
- face-detection-retail-0004
- human-pose-estimation-0001
- mobilenet-ssd
- openpose2
- pedestrian-detection-adas-0002
- person-detection-retail-0013
- person-vehicle-bike-detection-crossroad-1016
- road-segmentation-adas-0001
- tiny-yolo-v3
- vehicle-detection-adas-0002
- vehicle-license-plate-detection-barrier-0106
- yolo-v3
Videos of the the default and face detection models.
1.The nnInput window on the left displays the actual image input to the model
2. The color window on the right displays the image captured by the actual camera
python3 depthai_demo.py
python3 depthai_demo.py -gt cv -cnn face-detection-retail-0004
Summary
The PiNSIGHT works well. I have used Luxonis products in the past and the documentation is very good. The aluminum faceplate runs noticeably warm when inferencing but the Pi CPU stays around 55 C. When using the RPi AI Kit the Pi CPU runs around 72 C when inferencing. The Hailo-8L should outperform the Myriad X by a large margin but I'll need to figure out how to benchmark that.
I'll probably end up using the PiNSIGHT in integrated camera applications and dedicate the RPi AI Kit for IP camera stream processing.
