v0.5.3 is out. Many bug fixes thanks to your feedback, no new features (the “big refactor” arrives with v0.6.0).
Bug fixes
Exports now follow the source frame rate. 60fps footage no longer comes out at 30fps.
Mismatched camera frame rates now give a clear error instead of silently desynced output. Proper support planned.
AI detection no longer hangs the export when the ONNX Runtime library is missing or incompatible.
Audio carried across all chained segments, and seeking works across segments. Thanks @uSpike
Selecting files from “Recent” now actually loads them.
Several VRAM fixes for lookahead exports. Small cards should hit far fewer “not enough VRAM” walls.
Bug reports now keep the newest log lines. Old reports often got cut off.
AI on Windows and Linux
Windows: When building Reco locally, DirectML was silently not compiled in, so AI ran on CPU. It now runs on your GPU. Thanks @RufanMelfor
Windows: fixed a startup crash with broken Vulkan loaders. Thanks @RufanMelfor
Linux/macOS CLI packages now bundle ONNX Runtime, so AI works out of the box on prebuilt builds.
Plus what the two betas shipped: lens correction and seam blend persist in calibration, export failure reasons in the status bar, preview frees its GPU memory during export, multi-clip Windows export fixes, VRAM risk on the export slider, drag-to-reorder segments, NVENC bitrate scaling with resolution, VAAPI upload and quieter logging (thanks @uSpike), and reco-obs building on Windows (thanks @RufanMelfor).
I would like to thank y’all again for the feedback. Since v0.5.2: 140 commits, 84 merged PRs, 33 issues reported by you, 14 already closed, code from 3 contributors.
Keep up the good work Mohamed
Really look forward to trying this out.
Would tracking of things like distance covered, top speeds, heat maps for players be in your scope (if it’s even feasible) some time further down the line? I know everyone has slightly different things they want out of projects like this but I just wanted to ask because for me that would be the holy grail.
I think it would add so much value finally being able to scratch that Strava itch from football games. On top of the obvious benefits it would add to post game analysis.
The analytics like Veo is possible but I feel the focus should be nailing the stitching and broadcast crop flows first. I believe live streaming with a local processing device is next. Need to find someone who has done some work in this area before to help contribute.
Regarding analytics, the gap today is the AI. Like I mentioned in A community dataset to make tracking better, what do you think?, the gap today is in the ball and player tracking model. If we get that right, the analytics you are mentioning will be much more accessible.
Regardin @zzz’s comment, I agree that the stitching still has flaws. So does many core things, and they are my main priority now.
I waited to answer to your comments because I was excited to match it with the Reco Kit announcement Introducing the Reco Kit camera, because it achieves exactly what you have said: livestreaming with local processing device.
And what is even more exciting is that we managed to make Reco run on a much cheaper device (than Nvidia Jetson), at its full quality, locally, livestreaming to Youtube, which is what we are preparing with the Reco One.
