Sideway rotation and AI detection

[user_926183813247214_926183813247214]

Hello,

Does the AI detection depend on the angle of camera being straight (or at least in a landscape aspect ratio or similar)? I have bought and set up 3x 81MAs, and it makes more sense for the area I have to cover to have them in portrait mode (turned at around 90°)...

The cameras don't seem to reliably detect people in infrared mode, even when the motion sensitivity is set to 100 and the people smart detection is set to 100... Could it be because the cameras are rotated, and the AI cannot perceive sideways well?

It would be nice if Reolink made it so we could rotate the streams to be in portrait mode, or at least, if it isn't already the case, train their AI so that it detects people even when rotated in infrared conditions well... Otherwise I think the solution would be to re-encode each camera stream so that ffmpeg rotates it correctly... With a 4 camera setup, re-encoding seems like a lot of work, especially if I have to run another AI solution to detect infrared properly?

[user_920006777839800_920006777839800]

Yes, AI detection can be influenced by the orientation of the camera. Many AI models for object detection, including those used in security cameras, are trained on datasets where the objects (like people) are typically oriented upright in a landscape aspect ratio. When the camera is rotated, the AI might struggle to recognize objects correctly because the orientation of the features it was trained to detect has changed.

space waves

[user_943277391896724_943277391896724]

Here are a few suggestions to consider:

1. Camera Orientation: If possible, try to reposition the cameras back to a landscape orientation to see if detection improves.
2. Firmware Updates: Check if there are any firmware updates available for your cameras that might improve their performance or add features like stream rotation.
3. Feedback to Manufacturer: Consider reaching out to Reolink to suggest improvements for AI detection in rotated modes. User feedback can often influence future updates.
4. Re-encoding Streams: If rotation is necessary, using FFmpeg to re-encode the streams may be a viable solution, though it could be labor-intensive. Automating this process with scripts could help ease the workload.
5. Alternative AI Solutions: If you find the built-in detection insufficient, exploring third-party AI solutions that can process the video streams might be worth considering.

Ultimately, testing different configurations will help you find the best setup for your needs.

head basketball