This project is an ANN model for estimating single person poses from RGB image inputs, inspired by the model in the paper Simple Baselines for Human Pose Estimation and Tracking. The model features a ResNet18 backbone from the torchvision.models Python module (pre-trained), with the last two layers (Average pooling and FC layer) removed. Connected to this backbone is then 3 transposed convolutional (/deconvolutional) layers, which upscale the output feature maps of the ResNet18, with the output of the model being a heatmap for joint-location probabilities, with one channel for each joint (16 total).
The dataset used for training is the LIP dataset, which is a modification of the typical COCO dataset whereby, for the most part, each image has been cropped to contain only a single person. Hence, if one desires to train this model themselves, it is recommended to use the LIP dataset, as the code has been authored to assume that this dataset is the one being used. If one wants to use a different dataset, please ensure it is a single person dataset, and that you modify the code accordingly.
NOTE: the dataset images have not been included in the repository as it would needlessly clog it, hence if the LIP dataset is to be used for training, the training / validation image sets need to be downloaded from here. Then, put the validation image set in the res/val/img folder, and the training image set in the res/train/img folder. The corresponding annotations for the LIP dataset are already included in their respective res folders.
