Skip to content

Web tools

Two browser tools accompany the pipeline on the project page. Both run entirely in the browser: nothing is uploaded and no server is involved.

Rig checker

Drop a video. The page reads resolution, frame rate, codec and bitrate from the MP4 container without decoding it (the moov index is found wherever the camera put it, so a 130 MB clip is read in milliseconds), plays a preview, and shows a frame for labelling with zoom (fit, 2×, 4×) and a 4× loupe under the cursor so wall edges can be clicked to the pixel. The demo is a 2-second excerpt of a T-maze trial at the recording's native 3840 × 2160, so its scale reads as on the real rig. Then:

  1. Say what is in the video: T-maze, another maze or corridor, open field, or treadmill, plus whether the paws are visible, the whole arena is in frame and the camera is perpendicular. The gait module needs visible paws (a camera below a transparent floor, or a side view); only the T geometry enables the decision module; a treadmill needs the belt speed and direction.
  2. Mark the scale: two clicks on a known distance give px/cm.
  3. Mark the mouse: two clicks from snout to tail base cross-check the scale against mouse size (adult mice measure about 6.3 to 7.5 cm in STRIDE's cohorts). A value far outside that range means the scale is wrong or the animal is not a mouse. 3b. Mark a straight edge (4 or more points along a maze wall or a board row, with the loupe): the page fits a line and applies the pipeline's stage-1 rule, more than 2 px RMS bowing means undistort before pose inference. A failing check adds stride check-distortion to the recipe. This is the browser stand-in for the ChArUco-based check; the pipeline's board check remains the reference, since it also tests corner spacing and reprojection error.
  4. Read the verdict and the recipe: a per-module verdict against the recording requirements, what to change, and a generated config.yaml batch block and command list with your fps and scale filled in, the stages to skip for your setup, a note on which pose model to use, and links to the matching docs pages.

If the browser cannot decode the codec (HEVC from DJI cameras, for example), export one frame and drop it in for the scale step:

ffmpeg -ss 5 -i clip.MP4 -frames:v 1 frame.png

The checker scores the recording, not whether the trained pose models transfer to your rig.

STRIDE viewer

Load a .slp pose file and a .rois.yml region file, optionally with the video. The viewer parses the pose file with the lab's sleap-io.js, derives px/cm from segment2 exactly as the pipeline does, and runs a JavaScript port of the stride detector with the pipeline's default thresholds. It shows the regions, the tail-base trajectory and the current pose on the frame, a swing-and-stride timeline, summary tiles, and a per-stride table with CSV export. With a video loaded it plays the clip with the pose, timeline cursor and table row following; pause, scrub, or click a stride to stop on it.

The port reproduces the pipeline's stride boundaries on the bundled demo trial, and its stage-6 filters (confidence, edge strides, turning, backward strides) reproduce the pipeline's filtered set on that trial (11 of 18 strides). Each filter can be switched off, removed strides can be shown, and a speed band restricts the analysed strides to a range of walking speeds. Two clips can be loaded side by side (for example a wild-type and a tau trial) with a table of medians and their difference, best read at matched speed. For treadmill recordings, enter the belt speed and the direction the belt surface moves in the frame: keypoints are moved into the ground frame before detection.

One known difference: the forward-direction filter uses the clip's own median stride length where the pipeline uses the batch median. The pipeline's gait_per_stride_filtered.csv remains the authoritative table.