Contrast Studio
Interactive multimedia

RoadMap

Fi3lD 5/6/2019 road map

Road Map 

Better ID with existing bib reads:

Short-term:

Hungarian algorithm to handle association of detected bibs between frames.

Levenshtein distance applied to match partial or mis-read bib numbers

Medium-term:

Face detection and features from facial recognition for inter-frame matching.

Long-term:

Human pose estimation for inter-frame matching


Improved position tracking, robustness to occlusion:

Short-term:

Gaussian process or kalman filter to model motion

Long-term:

CNN to estimate 3d model of walkable path,

-project image-space detections into 3d,

-estimators that use linear velocity assumptions will be more reliable


Generally-improved runner ID:

Short-term:

Improved digit detector

Medium-term:

Improved OCR methods to read digits

Long-term:

Improved OCR to read entire bib, including names and other bib features

Facial recognition to relax bib-dependence


Processing time:

Short-term:

Convert object detectors to ONNX format to remove framework dependence

Remove unnecessary intermediate saving of images to disk

Medium-term:

Quantize object detectors for faster inference speed

Long-term:

Parallelize position tracking and sub-clip generation for scalability

Other Improvements:

Medium-term:

Automatically combine sub-clips of same person from multiple source videos

Long-term:

Generate runner statistics such as:

Real-world velocity

Steps per minute

Exhaustion level