Our modification of the stacked-hourglass architecture for car keypoint localization
Based on the architecture proposed in the ECCV 2016 paper [Stacked Hourglass Networks for Human Pose Estimation] (https://arxiv.org/abs/1603.06937)
Currently, the repository contains code for testing the network for car keypoint localization. Training code will be updated.
The code is written in torch, heavily borrows from https://github.com/anewell/pose-hg-train and has been modified to work on multi-GPU configurations.
To get started, run the prediction code in predictKpsLight.lua by executing
th predictKpsLight.lua
This script expects an input .txt file containing data in a format similar to that in testInstances.txt.
Specifically, each line contains one input instance in the following format
/abs/path/to/the/image.png x y w h
Here, x y w h are the (x,y) coordinates of the top left corner, width, and height respectively of the bounding box containing a car in the original image dimensions. The script takes care of the necessary cropping, scaling, transformations.
The script writes output (detected keypoints and confidence scores) to results.txt. These keypoints are according to the 64 x 64 cropped image size. These will have to be scaled to the original image size as per your requirements.
The trained model is available here https://www.dropbox.com/s/v4f770ebdenia0g/hg-binary-trained-singleGPU.t7?dl=0