Video Analysis for Logo Analysis using Deep Learning based Vision Models
The Challenge
Our client, a startup in the space of sports analytics, wanted to build a tool to provide quantitative feedback to marketers on the exposure they receive in return for their marketing spend at sporting events and its associated Video coverage. They also wanted to explore ways in which they could create clips to share on social media to get more “eye balls” on their brand.
Hypothesis
- Product Placement and advertising at sporting events captured on video coverage can be extracted, with accuracy from video.
- Video clips that can attract attention on social media can be detected.
- Viewership numbers can be merged with this extracted data to quantify return on spend.
Execution
- Pretrained models for object detection were used as a baseline.
- Further labelling of logos that were not being picked up by the Pretrained models was done manually.
- Baseline model was Fine-tuned for object detection using the labelled data.
- Models were also trained using the audio signal to curate clips of “key moments”.
- Compute the value of brand/product visibility based on viewership data for the event.
Outcomes
- The object detection was able to identify logos even with substantial obfuscation.
- A logo classifier assigned the logos to brands.
- Audio signal from videos could identify interesting “key moments”.
- Delivered a Quantitative measure of value per dollar spent.
- Logo Identification for known logos.
- Audio signals provide a first cut on “key moments” in a game. This is work in progress.
- Classification of logos into known brands and computation of value per dollar.
Project Highlights
95%
of brand impressions captured for increased ROI reporting.