YORU Tracker Identities and trajectories for YORU detections

Step4. Track All Videos and Analyse

Track every video

In the window

  1. Press Batch Tracking (or Go → Batch Tracking).
  2. Video folder → D:\experiment\videos.
  3. Detector model → the same model as before. Conf. as before.
  4. Settings… → Load… → two_flies.yaml → Apply.
  5. Output folder → D:\experiment\results.
  6. Press Run batch, and watch the Files table. A file that fails is marked failed with its error, and the batch goes on; the Failure report lists every failure at the end.

Or on the command line

uv run yoru-tracker batch D:\experiment\videos --model D:\models\best.pt --config D:\experiment\two_flies.yaml --out D:\experiment\results

Either way, each video gets three files:

D:\experiment\results\
├── pair01_tracks.csv        one row per fly per frame
├── pair01_tracks.json       settings, model, versions: how to reproduce it
├── pair01_detections.csv    the detector's output, to re-track without the model
├── pair02_tracks.csv
└── ...

Analyse the tracks

The tracks CSV is a plain table; its columns are described in Output Files. pandas and matplotlib are already in the YORU Tracker environment, so a script can be run with uv run python analyse.py:

# analyse.py
import json

import matplotlib.pyplot as plt
import pandas as pd

name = r"D:\experiment\results\pair01"
tracks = pd.read_csv(f"{name}_tracks.csv")
fps = json.load(open(f"{name}_tracks.json", encoding="utf-8"))["source"]["fps"]

# Position of each fly, one column per track ID
cx = tracks.pivot(index="frame_id", columns="track_id", values="cx")
cy = tracks.pivot(index="frame_id", columns="track_id", values="cy")

# Distance between the two flies, in pixels, per frame
distance = ((cx[0] - cx[1]) ** 2 + (cy[0] - cy[1]) ** 2) ** 0.5

# Speed of each fly, in pixels per second
tracks["speed"] = (tracks.vx ** 2 + tracks.vy ** 2) ** 0.5 * fps
print(tracks.groupby("track_id").speed.describe())

# Time each fly spent in each class (e.g. solo / copulation), in seconds
print(tracks.groupby(["track_id", "class_name"]).size().unstack(fill_value=0) / fps)

# Trajectories
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(11, 4))
for track_id, t in tracks.groupby("track_id"):
    ax1.plot(t.cx, t.cy, lw=0.7, label=f"ID {track_id}")
ax1.invert_yaxis()                  # image coordinates: y grows downwards
ax1.set_aspect("equal")
ax1.legend()
ax2.plot(distance.index / fps, distance)
ax2.set_xlabel("time (s)")
ax2.set_ylabel("distance between flies (px)")
plt.tight_layout()
plt.show()

Frames without a row. By default, a fly that was not detected in a frame has no row in that frame, so its position is NaN there after pivot. To have a predicted position instead, export with Include predicted rows of lost tracks (--include-predicted); those rows have predicted = 1.

Changing your mind later

Because the detections were saved, the whole experiment can be tracked again with different settings — in seconds per video, without the model or a GPU:

uv run yoru-tracker track D:\experiment\results\pair01_detections.csv --config new_settings.yaml --out D:\experiment\results_v2

The *_tracks.json beside each result records the exact settings it was made with.