Hi everyone,
So I got a question elsewhere about using temporal information in neural networks and I figured it may be of interest to post here as well:
Temporal information is very important in a lot of perception, whether it's integrating LSTM w/ YOLO, or using time to better estimate depth.
On the depth part, see here:
Which builds off of work from Google, which took advantage of time in the training data, but not in the network itself (which is the above improvement):
https://ai.googleblog.com/2019/05/moving-camera-moving-people-deep.html
On YOLO with an LSTM (so to take advantage of time), see here:
https://github.com/opencv/opencv/issues/15033
I think there's newer/better than that. AlexeyAB shared that back in July 2019... so that's from a LONG time ago in computer vision/AI rates.
We haven't investigated any of these on the platform... there's so much core platform work to be done, but many of them are probably runnable on the platform with some work/optimization.
We are looking forward to investigating something like these on DepthAI though, it could help with both object detection accuracy (including the bonus of having an integrated tracker) and help for better depth performance.
Cheers,
Brandon & The Luxonis Team
