Limitations - prnvkhndlwl/Kinetrace GitHub Wiki

  • Forward-only tracking (every point model runs online, forward in time): to fix the past, scrub back, correct, and re-track forward from there.
  • Regions can be circles, rectangles or polygons (Point β–Ύ); their members are sampled inside the outline. The animal silhouette covers the deformable case.
  • Several animals, no automatic re-identification. Each animal is its own layer in LAYERS, outlined and tracked in one pass, but identities come only from your clicks: when an outline jumps to a look-alike, that animal's run ends there (the others go on) rather than being matched back automatically. Importing multi-animal track files is not supported yet (exporting is).
  • Variable-frame-rate video (some phone/screen recordings): frame numbers become ambiguous. The app detects and warns; re-encode first: ffmpeg -i in.mp4 -vsync cfr -r 30 -c:v libx264 -crf 18 out.mp4
  • A file that states no usable frame rate: the app measures it from the frames' timestamps (or, failing that, assumes 30 fps) and says so when the video opens. Check that number, because every time, speed and camera sync uses it. High-speed rates up to 100 000 fps are read from the file.
  • A damaged frame stops tracking there ("Tracking stopped: frame N of the video could not be decoded") and keeps everything before it; scrubbing onto it shows a blank frame and says so (Shift+C retries).
  • Tracking speed depends on the point model and the picture size (measured on an NVIDIA GPU with refinement: β‰ˆ40 fps at 1080p with CoTracker3; at 4K β‰ˆ14 fps with AllTracker, the default, and 32–42 fps with CoTracker3). A full 40k-frame 4K pass with AllTracker is roughly an hour β€” pause/resume and autosave make that practical.
  • Without an NVIDIA GPU (a PC with no graphics card, or a Mac, which uses its Metal GPU) everything runs, several times slower: AllTracker on the CPU measured ~1.5 frames/s at 640Γ—480 and ~1 frame/s on a full 4K frame, so a long 4K video is an overnight job. SAM 3D Body is switched off there (Meta's code is CUDA-only; the Body dialog greys it out and offers the 2D model). On computers with little memory AllTracker works on a smaller picture (the system check names the size; a 4K full-frame window took 18 GB at the full size). Help β†’ System Check… lists, feature by feature, what this computer runs.
  • Intel Macs cannot run Kinetrace (PyTorch no longer builds for them).
  • Motion sync cannot see a small, far-away mover. 3D β†’ Sync β†’ Sync Cameras β†’ Motion compares how much each picture changes, on a small copy of the picture: a mover under about 1 % of it (on a 2704-pixel GoPro frame, roughly 234 Γ— 234 px or less, such as a wand tip far from the camera) leaves no trace, and the result reads unclear. Sync those cameras by sound instead (a clap or a beep that every camera hears), or wave something large in front of every camera at the start.
  • The wand's timing check finds whole frames only. It suggests an offset only when every pair of cameras agrees, and only for a camera that shares at least 30 wand frames with another; a fraction of a frame is named and left to Refine offsets with the wand. A wand always carried the same way at the same speed cannot time a camera (any offset looks alike): wave it in every direction.
  • One offset per camera. A camera whose clock runs at a slightly different rate (a drift) or that dropped a frame cannot be fitted by one offset; the timing checks say so but do not correct frame rates. The refine's verdicts were tuned on a synthetic rig: on a real rig they can be strict (rows read ok that the other cameras' prediction says improved) - nothing is applied unless you tick it.

← Performance (4K) Β |Β  Home Β |Β  Testing β†’