Training AI model on video data processed by Beamr’s content-adaptive technology made the model more resilient to compression, by lowering depth estimation error on safety-critical road users, including pedestrians and motorcyclists, by 30.7% 

Beamr Imaging Ltd. (NASDAQ: BMR), a leader in video optimization technology and solutions, released research demonstrating that machine vision models fine-tuned on video compressed by Beamr’s patented Content-Adaptive Bitrate (CABR) technology are more resilient than models trained on uncompressed data, while reducing the video data volumes that machine vision development depends on.

Machine vision teams handling petabyte-scale video data for autonomous vehicles (AV) and other video AI applications typically consider compression as a process for managing this scale. The findings reframe adaptive compression as an asset that strengthens AI model resilience, with the advantages of reducing storage and networking costs and infrastructure. This research extends Beamr’s ML-Safe benchmarks, validating a potential performance asset for AI models trained across machine vision applications.

Marketing Technology News: MarTech Interview with Stephen Howard-Sarin, MD of Retail Media, Americas @ Criteo

The research evaluated Depth Anything V2, a state-of-the-art monocular depth estimation model. The model was fine-tuned on AV video data compressed with Beamr’s technology that delivered 35.2% file-size reduction relative to baseline compression. The fine-tuned model demonstrated 30.7% reduction in depth estimation error on vulnerable road users, including pedestrians and motorcyclists, and 16.0% aggregate reduction across all object classes. Full methodology and results are available in the blog post.

“This research shows that compressed video data can produce models that are more robust, not less,” said Dani Megrelishvili, Beamr CPO. “That points to a different role for compression in our customers’ pipelines, from a cost they tolerate to a tool they deploy.”

Marketing Technology News: From MarTech Stack to MarTech Fabric: Weaving Brand, Content, and Conversion Into One Thread

“Machine vision teams have faced a structural trade-off: compress video data to manage scale, or face the escalating costs and infrastructure challenges of running AI models without compression,” said Ronen Nissim, ML Lead at Beamr. “Our research suggests this trade-off is more flexible than the industry may have assumed. By using compressed footage as augmentation during fine-tuning, we produced a model that performed better on the validation set than the equivalent model trained on uncompressed data.”

Beamr’s ML-safe benchmarks have previously validated content-adaptive compression across the AV development pipeline. The benchmarks demonstrated up to 50% file size reduction while preserving object detection accuracy at mean average precision of 0.96, with high fidelity across detection, localization, and confidence consistency. Subsequent testing for captioning workflows in world foundation model pipelines showed 41%–57% file size reduction with no measurable impact on the pipeline outputs.