Abstract

We present StreamDEQ, a method that aims to infer frame-wise representations on videos with minimal per-frame computation. Conventional deep networks perform feature extraction from scratch at each frame in the absence of ad-hoc solutions. We instead aim to build streaming recognition models that can natively exploit temporal smoothness between consecutive video frames. We observe that the recently emerging implicit layer models provide a convenient foundation to construct such models, as they define representations as the fixed points of shallow networks, which need to be estimated using iterative methods. Our main insight is to distribute the inference iterations over the temporal axis by using the most recent representation as a starting point at each frame. This scheme effectively recycles the recent inference computations and greatly reduces the needed processing time. Through extensive experimental analysis, we show that StreamDEQ is able to recover near-optimal representations within a few frames and maintain an up-to-date representation throughout the video duration. In this extended version, we further introduce two explicit variants, unrolled (UR-StreamDEQ) and stochastically unrolled (SUR-StreamDEQ) StreamDEQ, and validate our approach on video semantic segmentation, video object detection, and human pose estimation, where StreamDEQ achieves on-par accuracy with the baseline while being 2–4x faster.

Model Variants

StreamDEQ recycles the previous frame's representation as a warm start and performs only a few update steps per frame. We provide three streaming variants:

• IL-StreamDEQ

  The implicit-layer model, solving the fixed point per frame with a Broyden solver.

• UR-StreamDEQ

  Replaces the fixed-point solver with a fixed number of unrolled update steps.

• SUR-StreamDEQ

  Unrolls to a random depth during training for improved robustness across iteration budgets.

Video

Qualitative Results

The comparison below shows the baseline against the three streaming variants, each running only 2 update steps per frame. StreamDEQ progressively refines its representation as the video streams, closely matching the fully solved baseline at a fraction of the cost.

Baseline — 2 iterations/frame	IL-StreamDEQ — 2 iterations/frame
UR-StreamDEQ — 2 unroll steps/frame	SUR-StreamDEQ — 2 unroll steps/frame

Citation

If you find our work useful, please consider citing it:

@article{ertenli2026representation,
  title={Representation recycling for streaming video analysis},
  author={Ertenli, Can Ufuk and Cinbis, Ramazan Gokberk and Akbas, Emre},
  journal={Neurocomputing},
  pages={134366},
  year={2026},
  publisher={Elsevier}
}

@InProceedings{ertenli2022streaming,
  author="Ertenli, Can Ufuk and Akbas, Emre and Cinbis, Ramazan Gokberk",
  title="Streaming Multiscale Deep Equilibrium Models",
  booktitle="European Conference on Computer Vision (ECCV)",
  year="2022",
  pages="189--205",
  organization={Springer}
}