What it does

sense is an inference engine that runs pre-trained neural networks for human action recognition in real time on a CPU. It ships with two lightweight models (EfficientNet and MobileNet variants) trained on millions of videos, plus demo scripts for action recognition, fitness tracking, gesture control, calorie estimation, and repetition counting. There’s also a browser-based tool called SenseStudio to record videos and fine-tune a custom classifier on top of their feature extractors.

The interesting bit

The models collect four frames before each forward pass, then use strided temporal convolutions to maintain a larger receptive field—so they output predictions at 4 fps while keeping computational costs low. The README claims “as few as 2-5 videos per class” can yield “excellent performance” for custom classifiers, which is either impressive efficiency or optimistic marketing; the sources don’t clarify which.

Key highlights

• Two pre-trained backbone models, supposedly small enough for smooth real-time CPU inference
• 30 action classes, 80 fitness exercises, and 8 hand gestures recognized out of the box
• Calorie estimation via MET value prediction, with explicit caveat that accuracy is unverified
• SenseStudio browser tool for recording, annotating, and training custom classifiers
• iOS deployment path mentioned (details not shown in truncated README)

Caveats

• Pre-trained weights require separate account creation and evaluation license agreement at 20bn.com
• README installation steps only confirmed on Ubuntu 18.04/20.04 and macOS Catalina; Windows support unclear
• Calorie estimates “have not been verified in terms of accuracy” and are presented as relative metrics only

Verdict

Worth a look if you’re building fitness apps, gesture interfaces, or action-recognition prototypes and need something that runs locally without GPU dependency. Skip it if you need medically validated calorie data or can’t stomach the external licensing flow for weights.