Intro:
Crate an Image Processing tool, that given an image with a *single* person in it - finds that person's Contour (a shape that describes the edges around the person).
see contour-example.jpeg
Assumptions & constraints:
- Image resolution ranges between VGA to 720p
- Single person per image - since the image is already roughly cropped around the person
- Should work in various recording 'setups': different person's aspects / angles, clothing, person gender, environmental lighting conditions, locations, etc.
Input:
- A single image URI
- (if needed - can also be provided) A JSON structure with the pixel locations of the person's main joints (e.g. head, shoulders, elbows, wrists, hips, knees, ankles)
Output:
- Original image, with the body contour shape drawn over it.
- JSON structure containing an ordered list of 2D points (pixels), describing the person's contour
Threshold goals:
- Accuracy threshold: Mean IOU (Intersection Over Union) >= X (X to be provided as part of the Worker's offer), when compared to a ground truth JSON; Mean == averaged over 10 random images
- Analysis time threshold: up to 250ms per image, measured on a g3.4xlarge AWS EC2 instance
Additional deliverables:
1. A tool that will allow us to create the ground truth data: view an input image, manually draw the shape around the person, and click to export it into the same JSON format as above.
2. A tool that allow receiving X images with their 'ground truth' JSONs's, and then calculates & prints the IOU factor and analysis time for each image.
General Guidelines:
- Code will run on 'Deep Learning AMI (Ubuntu)' AWS instance
- It's requires to setup the EC2 instance, while using stable G.A. versions of the required installations
- Avoid using any licensed code that forbids changing, reusing, etc. for commercial purposes
- Version handling will be done in a Github repository; files that are too big for Github tracking, will be uploaded to a shared Dropbox folder, with full documentation
- Documentation will be written for: setup, configuration, and execution phases
- Code will be well structured, commented, and maintainable
About the recuiterMember since May 20, 2018 Arshad
from Tamil Nadu, India