Testing Gemini 3.5 Flash Lite for human detection in home surveillance (romanuk.org)

A custom home surveillance system utilizing Frigate and Casa Segura has recently tested the new Gemini 3.5 Flash Lite model for improved human detection, aiming to reduce false positives caused by pets. While the previous model, Gemini 3.1 Flash Lite, effectively managed detections without false alarms, the new model introduced a modest improvement in accuracy but at a 50% price increase. Despite the upgrade, Gemini 3.5 posted one false positive—a dog mistaken for a person—compared to 0 for the existing 3.1 model, raising concerns about its efficacy in practical scenarios. This development is significant for the AI/ML community as it highlights ongoing challenges in real-world applications of machine learning models, particularly in home security systems where reliability is paramount. The testing revealed similar performance in latency between both models; however, the solution to false positives did not lie in upgrading to the expensive option but rather implementing a secondary check within the same model. This suggests that iterative validation may be a more effective approach than merely upgrading to new models, prompting further discussion on model optimization versus pricing within the AI landscape.