1 of 7

CamSense AI:

Next-generation video recognition

VC Presentation (very cool)

2 of 7

Problem Statement

Existing video recognition technology doesn’t easily allow custom triggers based on content understanding of unattended recorded video

Eagle has landed

Flood threshold reached

wildfire precursor, public safety event, etc.

Other video feed (forest, street)

Eagle nest video feed

Local river

3 of 7

Solution

  • Use inference to gain content understanding of video clips with differences and measure similarity, saving time in data labeling, event identification and response

  • Cluster video clips based on embedding similarities

  • Human review needed to confirm labeling / identification

4 of 7

Solution wireframe

5 of 7

Solution explanation

  1. Submit clips (spliced video in 10-sec or other increment)
  2. 12 Labs discerns difference between any two video clips, one designated ‘reference’ and all others are used in comparison
    1. Output: Annotates the videos.
  3. Groq
    • Input: Annotations from 12 Labs
    • Output: Description of differences using sentence-transformers/all-mpnet-base-v2`
  4. If Then Logic:
    • Are the annotations similar enough to ignore?
    • How? Embedding distance measurement
    • If different enough log the clip, timestamp, and description…
  5. Aggregate all logs to make a final report

6 of 7

Funding Ask

  • $1050 in initial seed funding
  • Groq sponsorship through stylish gear and $250
  • $300 and 1 year of Claude Pro

7 of 7

Summary and Closing