Skip to content

v2.1.0: counts you can trust

Latest

Choose a tag to compare

@rampal-punia rampal-punia released this 07 Oct 18:29
· 10 commits to master since this release

Counts you can trust, an install that works everywhere, and better sample videos

Tracking fixes

  • Identical runs now give identical counts. Resetting a tracker used to stack Ultralytics' tracker callbacks, so every run after the first sent each frame through the tracker two, three, four times: on the same clip, 19 objects on the first run and 5 on every run after, and each run slower than the last.
  • Every visitor gets their own tracker. Tracking used one cached model shared by all browser sessions, so two people tracking at once fed one tracker (we measured 8 and 45 objects instead of 19 and 79).
  • Every tracker gets its own IDs. Ultralytics hands out track IDs from one counter per process and resets it whenever a tracker starts. Videos tracked side by side interleaved their IDs (churn 6.20 instead of 1.64), and a tracker starting mid-run could make another reissue an ID it already had. Multi-video now reports exactly what each video reports alone.
  • The webcam starts fresh IDs and counts on every START, and uses the same per-frame path as files, RTSP and YouTube.

New

  • ⏹ Stop button for playback.
  • The default image runs every task (it used to show a pre-rendered detection image under every task). Phone photos keep their orientation, and results survive reruns.
  • Nine new sample videos from Pexels, picked by running this app's own detector and tracker over 731 candidate clips. Credited in videos/ATTRIBUTION.md; the older clips are in videos/archive/.

Install

  • uv sync works on Linux, macOS and Windows (#19).
  • requirements.txt is now a portable list of direct dependencies (the old Linux + CUDA freeze failed on macOS and Windows, #18). The exact environment it was developed in is pinned in requirements-lock.txt.
  • packages.txt updated for Debian 12 (Streamlit Cloud). README: Python 3.10+, v2.1 features, and the licences (this code is Apache-2.0; the Ultralytics models are AGPL-3.0).

Tested headlessly through every task, source, tracker and ReID option, on both the pinned versions and today's latest (ultralytics 8.4.168, torch 2.14, opencv 5.0, streamlit 1.64).