Killing a C++ server to save a 2017 pose estimator in Keras

What it does This is a Keras reproduction of the 2017 CVPR paper Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. It converts the original authors’ pre-trained Caffe model into Keras format and provides training code against the COCO dataset. The main objective was dropping the original pipeline’s separate C++ rmpe_dataset_server in favor of a Python-native dataflow.

The interesting bit The author rebuilt the data pipeline using tensorpack so masks and augmentations happen on the fly without a separate C++ server. That fixed mask-assignment bugs and removed the need to pre-generate mostly-empty mask files.

Key highlights

• Converts the original Caffe weights to Keras; a pre-converted model is available via Dropbox.
• Replaces the old C++ rmpe_dataset_server with tensorpack dataflow for multi-process Python-native preprocessing.
• Generates masks on the fly instead of relying on pre-computed, often-empty mask files.
• Masks out unlabeled background people so the network doesn’t learn from ghost keypoints.
• The author reports training for roughly 10 hours on a single GTX 1070 (5 epochs / 25,000 iterations) to produce reasonable demo results.

Caveats

• The repository is explicitly deprecated; the author states it is incompatible with TensorFlow 2.0 and has moved active development to a fresh TensorFlow-centric repo.
• It is essentially glue and conversion code around a 2017 Caffe model, not a novel architecture.

Verdict Worth a look if you specifically need a Keras-native 2017 pose estimator and can tolerate legacy TensorFlow. Otherwise, follow the author’s own advice and jump to their newer TensorFlow 2.0 repository instead.