The background removal in .ai seems more jittery than in Vcam
under review
Jon Young
under review
Jon Young
Hi James, thanks for the feedback. If it's not too much trouble, it'd be great if you could send over two videos showing old vs. new - you can shoot over an email to joyo at splitmedialabs dot com. In any case, we'll do some further investigations our side and follow up here.
James Ellis
Jon Young
The VCam (when I could get it to work) legit looked like I was in the background. Maybe there was an antialiasing or just a subtle blur around the edges, but it was so nice.
Vcam.ai looks like Zoom three years ago.
Jon Young
James Ellis: We'll be putting together a build next week that ships with much more recent models. I'll follow up here once a download is available, it'd be interesting to see whether you see any improvement or not on that specific one.
Jon Young
James Ellis: Hi James, it took a little longer than hoped but we just released an updated version with newer models -- https://dashboard.vcam.ai/downloads.
Based on our testing, the models are now able to achieve a 97.5% IoU (Intersection over Union) -- we correctly segment all but 2.5% of the visible body. It'd be interesting to see how this fares for you compared to the previous version!
Obviously this is our biggest area of focus. We recently teamed up with Intel to improve this to 98.5% on their newer NPU chips (specifically designed to offload and handle AI based work), and over the coming months we'll be looking to collect more training data to cater to more scenarios, to improve all of our models.
For example, I can see in your screenshot that you are sat at an angle to the camera. Most of our training data is head on, meaning we can improve on this particular scenario. The model may also struggle if your lighting is weak (looks OK, though), or if your background is busy or has similar coloured objects to your skin or shirt. Again, with more data we will be able to better improve these scenarios. TLDR: onwards and upwards!