PIG-MENT

Given a video file as the sole input, the system returns three measurements without human intervention. Circling counts the animal's complete turns in each direction and detects whether it consistently favors one side—the most direct neurological indicator following intervention. The occupancy map shows where the animal spends its time within the pen, in actual centimeters, and how much time in each area of ​​interest. Rest time measures what fraction of time the animal remains motionless and how that time is distributed. Each video also includes its own quality indicators and control figures, ensuring that no single metric is ever presented in isolation.

The animal is located by its warm color, using segmentation, contours, and image timing. This decision has very positive practical consequences: not a single video needed to be tagged, no GPU is required, the infrastructure cost is low and predictable, and each result can be explained by the algorithm. Furthermore, it resolves the rest-by-construction problem: the animal is pink whether it is still or moving.

Each camera is calibrated once using a calibration pattern to correct lens distortion, and each recording automatically detects the four corners of the corral floor. With the actual corral dimensions, every measurement is expressed on the same plane in centimeters. Two different corrals or two camera positions produce comparable numbers, and the camera height isn't even configured: it's automatically determined from the validated geometry.

The video analysis, which is the most resource-intensive part, is performed only once and records the animal's position, area, and orientation in each frame. Recalculating a metric, correcting the scale, or recomputing the zones works on this data and takes only seconds. When the disk is full, the oldest footage can be deleted without loss: the results remain reproducible.

The web application guides the entire workflow: uploading the clip (directly to storage, bypassing the server, for files of several gigabytes), describing it, validating the camera's position against a reference image, processing it in the background, and reading the results. The recording date is read from the clock printed on the image itself, as the camera does not write it to the file. Phase, days since the intervention, and time slot are always derived from two dates, so correcting one corrects all pages.

The scientific value lies not in the video itself, but in the comparison. The animal's evolution is analyzed by combining recordings from sessions on a real-time timeline, starting from the intervention, and showing the change in each metric relative to its pre-operative state. The comparison tool compares groups of animals with a reference group and calculates the difference of differences, answering the study's question: Did the treatment alter the animal's evolution? Evidence filters remove excessively short clips from the median and alert you on screen. The tool presents the data; the researcher draws the conclusions.

Each export is a single file containing fourteen tables linked by identifiers, in Excel, CSV, and Parquet formats, plus the figures, trajectories, and a README file generated from the code itself, defining each column, each metric, and the rules used for calculation. It is possible to hide the treatment group from the results so researchers can work without being influenced by them.

Django and PostgreSQL, background processes running on two queues so interaction never waits behind a batch of videos, S3-compatible object storage, Dockerized infrastructure, automated deployment, error tracking, and a suite of synthetic trajectory tests that define the algorithm's behavior. The same discipline that Mecexis applies to every project, applied to a scientific measuring instrument.