3d-pose-baseline

Why we included this project

This is the code behind the ICCV 2017 paper that made the case for a simple baseline in 3D human pose estimation. The model takes 2D joint coordinates and lifts them into 3D positions, and it is deliberately compact so you can read the whole thing instead of wrestling with a large framework. Training on Human3.6M is straightforward: download the data, run the training script, and you can visualize results within a single epoch. That clarity makes it useful both as a benchmark for comparing newer pose-estimation work and as a teaching example of the regression-style approach later systems build on. One caveat: it relies on TensorFlow 1.x APIs and expects ground-truth 2D detections, so treat it as a research reference rather than a production pipeline.