Video analytics can provide benefits to manufacturing by enhancing revenue, reducing costs, improving quality and capacity utilization, and ensuring compliance. It has applications in process monitoring, intrusion detection, and quality checking. Challenges include handling large video files, network bandwidth limitations, specialized hardware needs, and extracting business value from raw data. Effective strategies involve moving processing to edge devices, only transferring necessary data, designing for repeatability, implementing feedback loops, starting with simple analyses, leveraging video properties like static frames, and exploring self-supervised and semi-supervised learning techniques. Future work may include federated learning and new model architectures for edge devices.
4. Caution!!
Buzz words and fancy numbers ahead!
● Industry 4.0
● Manufacturing to go from ~16% of GDP in 2019 to ~25% in
2025
● Industry 4.0 market cap to go from $66.10 Billion in 2017 to
$155.30 Billion in 2024
5. Benefits
1. Enhances top-line
2. OPEX reduction
3. Increased capacity
utilisation
4. Improved quality
5. Adherence to international
standards
15. Challenges
● Heavy video files
● Limited network bandwidth
● Need for specialized hardware
● Extracting business value from raw data
● Human stupidity
17. General gyaan
● Move heavy lifting to edge
● Transfer data only when needed and only what is needed
● Design for repeatability
○ Find common patterns
● Plan for ‘fool-proofing’
● Feedback loops
○ Data shift happens
● Pick your battles wisely
○ Have milestones
21. Video data specific gyaan
● Make use of static nature of videos
○ No specialized processing needed
○ Even less data suffices
○ But be wary of rare events
● Not every frame is useful
○ Avoid processing if you can
● Abundance of raw video data
22. Video data specific gyaan
● Abundance of raw video data
○ Where’s my self-supervised learning algorithm
■ Process-of-interest
■ object-of-interest
■ Event-of-interest
○ Semi-supervised/weakly-supervised learning
■ Student-Teacher paradigm
24. Video data specific gyaan
● Supervised learning
○ Use augmentations wisely
■ Not every augmentation is useful
■ But may help create robust model
○ Feedback loops of uncertain data
○ Be innovative in annotating data
■ Use temporal information
■ Mask vs bbox
■ Not every sample is equally useful
25. Some exciting experiments in the pipeline
● Federated learning
○ Saves infra
○ Saves bandwidth
○ Better models
○ Privacy friendly
● Slow learning of temporally coherent matrices (unsupervised)
● Faster, lighter models capable of running in parallel on small
edge devices
26. The best way to predict the
future, is to create it.