Reda Tagmouti

From idea to decision

1. 01The hypothesis: use computer vision to detect exercise form and give feedback during the movement.
2. 02I first built for gyms, where cameras could monitor multiple users. I trained models for squats, push-ups, pull-ups and dips. The system worked, but three constraints emerged: the pipeline scaled to only around 10 simultaneous users, installing the required setup was too expensive for a gym, and the legal side was a blocker. Tracking each person individually on the gym's cameras without their consent wasn't possible under GDPR. In a place open to the public, people can't exercise their right to object, and the CNIL only accepts this for anonymised statistical uses.
3. 03I pivoted to mobile to reduce deployment friction. I took the prototype to Delta Festival (Marseille), where visitors could test live pull-up and dip counting using a camera and a laptop. The stand stayed busy throughout the weekend, with 2,000+ visitors at the stand over the weekend, giving me an initial signal of user interest.
4. 04I then built the backend for video streaming and rep detection, followed by a second prototype using the Gemini API to analyze frames at the key moment of each rep.
5. 05The market study changed the decision. Users were willing to pay around €5, while estimated operating costs were around €10 per user under the tested assumptions. I tried to close the gap by replacing the API with a cheaper CPU-based model, using the API only to collect training data. The resulting on-device pose accuracy wasn't reliable enough for real-time feedback. I stopped the project.

What I learned

Demand alone wasn't enough. The product needed to work technically, scale at an acceptable cost, and support a viable price point at the same time.

The annotation tooling

Labelling was the bottleneck of the project, and off-the-shelf platforms weren't good enough for this dataset — so I built the tool around it.

• 01Bootstrapped labels with YOLOE: I show it the object on an easy dataset, correct what it misses, train a first model, then use that model to re-annotate the rest.
• 02SAM fits every box to the object's real shape, so I no longer adjust bounding boxes by hand.
• 03Duplicate-annotation detection, and propagation of one label across dozens of frames when the object stays put and only the background changes.
• 04A stream watcher that crops detections straight into the dataset, removing most of the manual work of finding and cutting footage.
• 05Side-by-side model comparison, a review pass over the labelled set to catch errors, and a separate mode for classification datasets.

For the labelling loop I compared YOLOE against Grounding DINO, OWLv2 and LocateAnything — YOLOE Large came out ahead, and it anchors the loop. I run the same comparison across the models I train myself, to pick the one I re-annotate with.