Computer vision consultant
As a Computer Vision Consultant at OpenMotion, I was responsible for evaluating the feasibility of a specific computer vision project. I began by discussing the requirements and constraints with the project team to determine the best approach.
Next, I conducted extensive research on state-of-the-art computer vision technologies and analyzed how they could be applied to the project. Based on my research, I identified the most appropriate solution and created a detailed report explaining why it was the best fit for the project. The report was approximately 30 pages long and included a thorough analysis and recommendations.
Interns mentor
At OpenMotion, I had the opportunity to mentor and guide two interns who were exploring AI solutions and developing proof of concepts. As a mentor, I was responsible for providing support and guidance to ensure their success. Over the course of six months, I mentored them closely, providing regular feedback and guidance as needed. The first project was focused on machine learning, with the goal of using signals from a smart watch’s accelerometer and gyroscope to detect falls. The second project was centered around computer vision, with the goal of detecting and reading license plates using YOLOv4 and Tesseract.
Computer vision Engineer
OpenMotion is a small but thriving and innovative company aiming to add some artificial intelligence in their services. Nowadays, cleaners are subjected to some tedious process. OpenMotion already worked to ease these process by proposing an automated solution. Cleaners don't waste their time bothering with process aymore and companies still get their reports. However, there is still some process they want to improve to be fully automated. One of them is taking a picture before and after they emptied the bin. To get rid of this process, OpenMotion proposed to use smartglasses that would detect and take picture of bins for them.
First solution was to embbed a convolutional neural network on an Android phone. I worked alone on the computer vision part and shared the integration in Android. However, I did the AI class alone.
- EfficientDet Lite
- SSD Mobilenet v2 FPN Lite
- Data collection and labelisation
- Train models
- TFLite converter script
- Tool to evaluate TFLite models (precision, recall)
- Documentation
Second solution was to compute on a server and develop a client on Android. I did not train the model but I've set up the client to send images and receive them and compute them on server.
- YoloV4
- RTMP/RTSP remote computing
- Documentation
Sucessfully integrated on Android:
- 85% Precision
- 100% Recall
- Real-time: +15fps on a Honor View10
Image source can come from the camera or from an external device, such as smartglasses.
- Bordeaux • France
- (+33) 6 33 48 16 27
- julien.hongsavanh@gmail.com