Source-linked AI summary

Temporally Distributed Networks for Fast Video Semantic Segmentation

Ping Hu, Fabian Caba Heilbron, Oliver Wang, Zhe Lin, Stan Sclaroff, Federico Perazzi

arXiv:2004.01800v2cs.CVcs.LGcs.MMeess.IV

TL;DR

Video semantic segmentation requires strong features but must also meet low-latency demands, which deep per-frame computation makes difficult. TDNet distributes shallow sub-networks across sequential frames and uses attention propagation plus grouped distillation to reconstruct strong features efficiently. Across Cityscapes, CamVid, and NYUD-v2, it consistently outperforms previous methods in accuracy and efficiency.

  • Problem

    Deep feature extraction improves video segmentation accuracy but its computation cost makes high-accuracy, low-latency processing difficult.

  • Method

    TDNet distributes feature-group extraction across sequential frames, recomposes groups with attention propagation, and applies grouped knowledge distillation.

  • Results

    TDNet consistently outperforms previous video segmentation methods in both accuracy and efficiency across Cityscapes, CamVid, and NYUD-v2.

  • Takeaways & Limitations

    Temporal distribution lets TDNet retain high accuracy while significantly reducing video-frame processing latency.

Abstract

from arXiv · show

We present TDNet, a temporally distributed network designed for fast and accurate video semantic segmentation. We observe that features extracted from a certain high-level layer of a deep CNN can be approximated by composing features extracted from several shallower sub-networks. Leveraging the inherent temporal continuity in videos, we distribute these sub-networks over sequential frames. Therefore, at each time step, we only need to perform a lightweight computation to extract a sub-features group from a single sub-network. The full features used for segmentation are then recomposed by application of a novel attention propagation module that compensates for geometry deformation between frames. A grouped knowledge distillation loss is also introduced to further improve the representation power at both full and sub-feature levels. Experiments on Cityscapes, CamVid, and NYUD-v2 demonstrate that our method achieves state-of-the-art accuracy with significantly faster speed and lower latency.

1. Introduction

Video semantic segmentation needs strong features for accuracy, but deep networks impose high computation costs that hinder low-latency applications. TDNet distributes shallow feature extraction across frames, then recomposes feature groups with temporal attention and grouped distillation to improve the speed–accuracy trade-off.

  • Motivation: Deep networks provide strong features for accurate video segmentation but are costly for latency-sensitive applications.The paper highlights self-driving cars, robot sensing, and augmented reality as relevant settings.
  • Motivation: Applying a deep segmentation model independently to every frame fails to exploit temporal information and recomputes all features.
  • TDNet: TDNet divides deep-model features into groups approximated by distinct shallow sub-networks with comparable total model capacity.The sub-networks share an architecture scaled to 1/N of the original deep model’s size.
  • TDNet: Sequential circular assignment computes one new sub-feature group per frame while reusing complementary groups from earlier frames.This distributes computation over time and avoids recomputing all feature groups at every frame.
  • TDNet: Attention propagation compensates for motion-induced spatial misalignment when reassembling feature groups across frames.Grouped knowledge distillation further improves representation at both full-feature and sub-feature levels.
  • Reported contributions: TDNet reports more than 2× lower latency at comparable accuracy and better accuracy and latency than prior methods across Cityscapes, CamVid, and NYUD-v2.

2. Related Work

Prior video segmentation methods either add temporal layers to repeated deep inference or propagate keyframe features. These approaches trade accuracy against computation or face motion-induced spatial misalignment during feature reuse.

  • Image semantic segmentation: Image segmentation research has pursued either efficient models or high-quality models after deep learning enabled dense label prediction.
  • Video semantic segmentation: Temporal-context methods repeatedly apply a deep model and aggregate features, improving accuracy over single-frame approaches but adding per-frame computation.
  • Video semantic segmentation: Keyframe-based methods reduce computation by propagating and reusing high-level features, but motion can misalign pixel-level information over time.

3. Temporally Distributed Network

TDNet distributes feature extraction across sequential video frames, then recomposes complementary sub-features with attention propagation to produce segmentation features efficiently despite motion-induced misalignment.

  • 3.1. Distributed Networks: TDNet divides a deep model’s feature representation into sub-features approximated by distinct shallow sub-networks.The sub-networks share a shallow architecture sized to preserve similar total model capacity while separating feature subspaces.
  • 3.1. Distributed Networks: Sequential circular assignment computes one new sub-feature group per frame and reuses earlier groups to avoid redundant feature extraction.The full feature representation is formed by aggregating previously computed feature groups with the current one.
  • 3.2. Feature Aggregation: APM addresses motion-induced spatial misalignment by propagating attention between feature groups extracted at different time steps.Unlike optical-flow warping, the module is based on non-local attention extended for spatio-temporal variation.
  • 3.2. Feature Aggregation: For four sub-networks, each four-frame window covers all sub-networks, allowing frame t to merge its current map with maps from t-3, t-2, and t-1.The next frame reuses the corresponding three preceding sub-feature maps in a shifted temporal window.
  • 3.2. Feature Aggregation: Downsampling and local temporal propagation reduce attention cost while restricting comparisons to neighboring frames with smaller motion.The attention complexity is reduced by a factor of n^2, and experiments find n=4 effective for preserving spatial information while lowering computation.
  • 3.2. Feature Aggregation: Grouped knowledge distillation aligns both full student outputs and sub-network outputs with corresponding teacher representations.The grouped objective conditions the teacher output on each respective feature subspace.

4. Grouped Knowledge Distillation

Grouped knowledge distillation transfers knowledge from a strong deep teacher to TDNet at both the whole-model and sub-feature-group levels. This encourages complementary information across feature paths while supervising the student’s segmentation outputs.

  • Grouped distillation transfers knowledge at the subspace level so information from different feature paths becomes more complementary.
  • A deep baseline such as PSPNet101 serves as the teacher, while TDNet with m sub-networks serves as the student.
  • The loss aligns student and teacher output distributions both for the complete model and for individual sub-feature groups.
  • The teacher’s feature reduction layer is evenly divided into m independent sub-convolution groups to produce the teacher sub-feature groups.
  • The final loss combines ground-truth cross-entropy supervision, whole-model KL-divergence distillation, and feature-group KL-divergence distillation.The paper sets α and β to 0.5.

5. Experiments

Experiments on Cityscapes, Camvid, and NYUD-v2 show that TDNet maintains or improves segmentation accuracy while reducing computation, latency, and temporal-load imbalance. Ablations further support attention propagation, grouped distillation, distributed feature paths, and moderate attention downsampling.

  • Setup and overall results: TDNet achieves state-of-the-art accuracy with faster speed and lower, evenly distributed latency on Cityscapes, Camvid, and NYUD-v2.The evaluation uses PSPNet and BiseNet-based backbones and reports speed on a single Titan Xp with batch size one.
  • Cityscapes: TD2-PSP50 reduces Cityscapes per-frame time from 360ms to 178ms while improving accuracy relative to PSPNet101 with similar total parameters.TD4-PSP18 reduces latency further to one quarter, but its shallower subnetworks reduce performance relative to PSPNet101.
  • Cityscapes real-time models: TD2-Bise34 outperforms prior real-time methods at comparable speed and is roughly three times faster than similarly sized BiseNet∗101.TD4-Bise18 reaches nearly 50 FPS with lower accuracy, while both distributed models improve over their single-path baselines at similar time cost.
  • Camvid: TD2-PSP50 surpasses Netwarp by about 9% mIoU while being roughly four times faster on Camvid.Compared with similarly capacitated PSPNet101 baselines, it approximately halves computation with comparable accuracy; TD4-PSP18 halves latency again with an accuracy decrease.
  • NYUD-v2: On NYUD-v2, TD2-PSP50 outperforms STD2P in accuracy and speed, halves PSPNet101 latency while preserving accuracy, and improves over PSPNet18 by about 1.6% mIoU without increasing latency.TD4-PSP18 is more than 5× faster but has worse accuracy.
  • Method analysis: APM outperforms optical-flow warping and vanilla spatio-temporal attention in accuracy and latency, remaining robust with a six-frame sampling gap.Without motion compensation, simple feature addition has the worst accuracy; flow-based methods fail at the six-frame gap while APM drops only slightly.

6. Conclusion

TDNet combines temporally distributed feature computation with attention-based merging to improve video segmentation efficiency while retaining accuracy. Grouped knowledge distillation further boosts performance, and the method consistently outperforms prior approaches in accuracy and efficiency.

  • TDNet merges feature maps computed across different frames with attention propagation to improve processing latency while retaining high accuracy.The attention propagation module handles feature alignment when recombining temporally distributed features.
  • Grouped knowledge distillation further boosts TDNet’s performance.
  • TDNet consistently outperforms previous methods in both accuracy and efficiency.
Loading 2004.01800v2…