Streaming Learning Center Where Streaming Professionals Learn to Excel
Related Articles
I recently visited Florida Atlantic University’s Multimedia Lab to record the first real-time demonstration of Feature Coding for Machines (FCM), a new approach to video compression designed for AI and edge-based applications. The system, developed by OP Solutions in collaboration with FAU, demonstrates low-power, efficient operation. You can watch a video of the demonstration and its implications here, and it’s embedded below.
As video is increasingly deployed in critical systems such as autonomous vehicles, smart cities, surveillance, and drone swarms, traditional codecs optimized for human viewing can’t meet the bandwidth, latency, and power demands of machine-driven operations. FCM addresses this challenge by reengineering the process of encoding, processing, and interpreting video for machine-driven applications.
Traditional codecs like H.264/AVC, H.265/HEVC, and H.266/VVC attempt to reproduce the original image as faithfully as possible, pixel by pixel. That makes sense when people are watching the video, but not for machine consumption.
Machine-driven applications rely on patterns, shapes, and textures, known as features, to enable object classification and tracking. By optimizing compression to deliver only the features relevant to the task, technologies like FCM enable more bandwidth and energy-efficient operation.
Contents
FCM Demo
OP Solutions, working in collaboration with Florida Atlantic University, has developed the first real-time implementation of Feature Coding for Machines. The demonstration uses a robot dog as the edge device, capturing live video and transmitting only extracted features to a remote server. The server interprets these features, recognizes the human in the scene, and sends commands back to the dog to follow.
Because the transmitted stream contains only the data needed for object detection rather than the full image, it operates at a fraction of the bandwidth of traditional codecs. It also preserves privacy, since the stream reveals no identifiable visual information about the environment or the people within it.
Figure 1 illustrates this process. In the top-left panel, the edge device, in this case the robot dog, captures live video of a human in its field of view. A portion of the neural network runs locally on the device to extract visual features from that video. These extracted features, not the full image, are encoded by the FCM encoder and transmitted as a compact bitstream, shown in the center panel.
On the right, the server receives the bitstream, decodes it, and processes the information to detect the human, represented by the bounding box. This shows how FCM transmits only privacy-compliant feature-related data, reducing bandwidth while maintaining accurate detection.
Figure 2 compares the traditional video compression model with the FCM approach. In conventional systems, the complete image is sent from the edge device to the server, and all video processing occurs on the server after the full video stream is transmitted.
With FCM, AI processing is divided between the edge device and the server. The edge device performs the initial neural network operations, encodes the extracted features, and sends only this lightweight data for final analysis on the server. This division of processing reduces bandwidth use and power consumption while maintaining accuracy.
FCM Efficiency
Figure 3 quantifies the bitrate needed by different codecs to achieve the same level of object detection accuracy.
As the figure shows, FCM achieves similar accuracy at roughly 10% the bandwidth of H.264 and just less than 33% the bandwidth of VVC. This improvement directly impacts scalability. In environments with large numbers of connected devices, like fleets of autonomous vehicles, smart city sensors, or coordinated drone operations, every reduction in bitrate translates to lower network load, reduced latency, and greater reliability under constrained conditions.
Deployment Timeline
Figure 4 illustrates FCM’s standardization and deployment timeline, showing that both OP Solutions and FAU have been actively involved in these standardization efforts since their inception.
As you can see, Feature Coding for Machines is already well into the standardization process. The committee draft was scheduled for September 2025, followed by the draft international standard in October 2025 and a fully compliant product launch in July 2026. With real-time operation already demonstrated, FCM is well-positioned to become the foundation for how AI systems transmit and interpret video at scale.
OP Solutions and Florida Atlantic University have contributed several core tools that are now part of the official reference software and are authoring the FCM reference manual to be published by Springer. OP Solutions is already engaging potential partners to support and expand participation in these initiatives and is well-positioned to deliver technology that enables these operations.
Feature Coding for Machines represents a significant step forward in utilizing video for AI-driven applications. By reducing bandwidth, lowering power consumption, and improving the speed and accuracy of machine perception, FCM enables intelligent systems to see more, act faster, and operate efficiently at scale. The technology has already been proven in real-time and is progressing toward full standardization.
Watch the video to see the demonstration and learn more about how OP Solutions and Florida Atlantic University are advancing this new approach to video for machine-driven operations.
For more information, contact OP Solutions here.
Haha, Feature Coding for Machines sounds like the ultimate solution for my always-overwhelmed home Wi-Fi when the drone swarm decides its time for a synchronized interpretive dance. Who needs pixel-perfect cat videos when your robot dog can just beam over quick, something round and kinda fuzzy over there! FCM is the digital equivalent of sending just the punchline – brilliant for efficiency and, apparently, maintaining *some* semblance of privacy when the machine cant tell Grandma from a pizza box. Bring on the standardization, I say, just as long as it doesnt require retraining my AI to differentiate between intruder and that suspiciously large dust bunny.football.bros
Haha, Feature Coding for Machines sounds like the solution when your cat keeps watching videos meant for autonomous vehicles, but the bandwidth bill gets too high! FCM is like teaching your pet robot dog to bark human! instead of transmitting the whole embarrassing scene. Efficient, privacy-preserving, and probably less confusing than trying to decode why the dog keeps chasing squirrels in the transmitted video. Truly, the future of seeing without the full picture – just the neat little boxes saying Person Here!. Very clever, OP Solutions!the prophecy deltarune
Haha, so FCM finally gives AI the birds-eye view it craves without hogging the bandwidth! Forget sending pixel-perfect masterpieces; machines just want the whats where? lowdown. Its like teaching your robot dog to bark only when it spots treasure, rather than recite the entire encyclopaedia. Efficient, privacy-preserving, and sounds just about *perfect* for when AI wants to play hide-and-seek with objects instead of humans. Cant wait for July 2026 when my smart fridge finally tells me the milk is low without uploading my entire breakfast habits!vòng quay may mắn