By the Koshee Team • ISC West 2026 • The Venetian Expo, Las Vegas • March 2026
At ISC West 2026, our CSO James Edmondson, PhD, a longtime computer vision researcher at Vanderbilt and Carnegie Mellon, delivered a technically grounded session on the realities of achieving accurate object and human activity detection in security. Below are the highlights and key takeaways for security practitioners looking to move beyond the hype.
The Context: Learning from AI History
James opened with a 70-year history of machine learning, from the 1957 Mark I Perceptron to today’s transformers. He reminded the audience that “hype cycles” are not new; the AI winters of the ’80s and ’90s occurred because technology was often overpromised and under-delivered. At Koshee, we believe the industry must prioritize sustainable, proven results over marketing trends to avoid a similar pattern today.
The Foundation: Data Quality is Everything
A central theme of the session is that model performance is fundamentally determined by data quality. James discussed the limitations of using certain open-source datasets, such as MS COCO, as a primary baseline. These datasets can often be mislabeled or lack the specific context required for high-stakes security environments, such as stadium crowds or varying focal lengths.
James’s Recommendations:
- Invest in Custom Labeling: Use trained, consistent labeling rather than relying solely on automated tools.
- Establish Clear Policies: Ensure a consistent labeling scheme across your entire team.
- Test for Reliability: Instead of only looking for successful detections, rigorously test your models against “failure cases” to ensure they handle real-world complexity.
Balancing Performance: Frameworks and Deployment
James surveyed the current landscape of detection frameworks, noting that while Transformers are excellent for multimodal tasks, focused frameworks like YOLO are often more practical for real-time security needs. However, he emphasized that not all models are created equal and warned against choosing deployment methods solely based on cost.
One critical concern shared by our team is the risk of extreme quantization (such as INT8 for certain hardware). While it reduces hardware costs, it can introduce unacceptable accuracy losses for mission-critical applications. For high-stakes security, maintaining higher precision (like FP16) ensures that your system remains a trusted tool rather than a source of false alarms.
These are principles we build on every day at Koshee. If you’re working through similar challenges in your own deployments, we’d love to connect.
Watch James’s full ISC West session: HERE