FOUNDATIONS / ICLR 2017 + CVPR 2019
Learn the representation.
Learned compression links representation design to rate and quality, then extends the problem across successive video frames.[1][2]
RESEARCH & PERSPECTIVES
Selected research that informs how the industry thinks about visual data. Our perspective connects these ideas to the practical questions facing AI application builders.
FOUNDATIONS / ICLR 2017 + CVPR 2019
Learned compression links representation design to rate and quality, then extends the problem across successive video frames.[1][2]
DEPLOYMENT / CVPR 2025
DCVC-RT brings execution overhead into the discussion of practical neural codecs.[3]
MACHINE VISION / MPEG
Machine-oriented coding asks how compressed information supports the downstream task, alongside its bitrate.[4]
Independent research, interpreted by Deep Refactor. These references describe third-party work, not Deep Refactor products, partnerships or measured performance.
The opportunity is to make visual information more useful within real constraints. That requires more than a compression-ratio headline.
The question is not simply how to make a video file smaller. For an AI application, it is which information must survive, how quickly it must arrive and what resources the complete pipeline can afford. A design that suits an archive may be unsuitable for a live inspection system. This is an evaluation framework, rather than a claim that a single codec fits every setting.
Ballé, Laparra and Simoncelli's ICLR research optimizes learned image transforms around bitrate and reconstruction error. The underlying question carries into video: what quality is retained at a given data budget? Quality must be defined for the intended use; a smaller stream alone does not establish a better result.[1]
Successive frames contain related information. Conventional predictive pipelines exploit that relationship through motion prediction and residual coding. DVC demonstrates an end-to-end learned formulation in which these components are optimized together. Hybrid approaches and neural approaches therefore need comparison at the system level, rather than being treated as two completely unrelated ideas.[2]
DCVC-RT addresses operational costs such as memory access and repeated function calls, alongside computation. This makes deployment efficiency a design concern. Its reported performance remains specific to the authors' implementation and tests. For a prospective application, latency, hardware, rate control and interoperability should be assessed together.[3]
MPEG's Video Coding for Machines activity considers performance on machine tasks after decoding. That widens the evaluation lens beyond human viewing. Task-aware and semantic approaches should not be assumed to preserve every pixel, every future task or evidential detail. A representation optimized for one use may need further validation for another.[4]
Our view is that useful evaluation starts with representative footage and explicit acceptance criteria. Compare end-to-end cost, temporal stability, failure behaviour and the downstream result. Test difficult cases as well as typical scenes, and state what was measured. These questions guide a research conversation; they do not substitute for validated product performance or imply that Deep Refactor has implemented the cited systems. An initial pilot could compare a standards-based baseline, a learned codec and a task-oriented representation using the same footage and delivery constraints. The decision can then follow the application's priorities, including graceful fallback and the ability to revisit stored material when requirements change.
PRIMARY SOURCES / FURTHER READING
Authors' research publication; hosted by New York University · 2017
An image-coding foundation, not evidence of video-system performance or universal superiority.
Peer-reviewed conference paper · 2019
Reported comparisons depend on the paper's codecs, test conditions and quality metrics.
University of Science and Technology of China / Microsoft Research Asia · 2025
Laboratory speed and compression results are hardware- and test-dependent; no Deep Refactor benchmark is implied.
Moving Picture Experts Group (MPEG) · Ongoing activity; reviewed November 2026
A standardization direction, not a general guarantee of accuracy or a company endorsement.