Source-linked AI summary

Video Object Segmentation with Adaptive Feature Bank and Uncertain-Region Refinement

Yongqing Liang, Xin Li, Navid Jafari, Qin Chen

arXiv:2010.07958v1cs.CV

TL;DR

Semi-supervised matching-based VOS must organize a continuously growing feature bank while accurately segmenting uncertain regions. The paper introduces adaptive feature-bank updates and uncertain-region refinement with confidence loss and local fine-grained segmentation, and reports outperforming state-of-the-art methods on two large-scale benchmark datasets.

  • Problem

    Matching-based VOS lacks an effective way to organize continuously growing feature-bank information, while low-resolution decoding produces ambiguous boundaries that affect segmentation accuracy.

  • Method

    The framework dynamically merges new features and removes obsolete ones through an adaptive feature bank, while confidence loss and local refinement address uncertain regions.

  • Results

    The approach outperforms state-of-the-art methods on two large-scale benchmark datasets.

  • Takeaways & Limitations

    Adaptive feature organization and uncertain-region refinement support segmentation of long videos while maintaining low memory consumption.

Abstract

from arXiv · show

We propose a new matching-based framework for semi-supervised video object segmentation (VOS). Recently, state-of-the-art VOS performance has been achieved by matching-based algorithms, in which feature banks are created to store features for region matching and classification. However, how to effectively organize information in the continuously growing feature bank remains under-explored, and this leads to inefficient design of the bank. We introduce an adaptive feature bank update scheme to dynamically absorb new features and discard obsolete features. We also design a new confidence loss and a fine-grained segmentation module to enhance the segmentation accuracy in uncertain regions. On public benchmarks, our algorithm outperforms existing state-of-the-arts.

1 Introduction

Semi-supervised VOS must distinguish target objects from background and preserve clear boundaries across subsequent frames. This framework addresses long-video memory management with an adaptive feature bank and improves ambiguous boundaries through uncertain-region refinement.

  • 1 Introduction: Semi-supervised VOS uses a first-frame annotation to segment the target object in subsequent frames.Performance depends on distinguishing object regions from background and clearly segmenting object boundaries.
  • 1 Introduction: Matching-based VOS cannot memorize all previous-frame information because memory is limited and not all frames are necessary.Existing methods may use only first and latest frames, risking missed key frames or out-of-memory failures as videos lengthen.
  • 1 Introduction: The adaptive feature bank absorbs new features by weighted averaging and discards obsolete features using a least-frequently-used index.This organization is designed to memorize characteristics of multiple objects while maintaining low memory consumption in long videos.
  • 1 Introduction: Uncertain-region refinement combines a classification confidence loss with local fine-grained segmentation to improve ambiguous boundary regions.The approach targets errors that can accumulate when boundary masks are difficult to classify and become vague after iterative upscaling.
  • 1 Introduction: The paper presents the adaptive feature bank, confidence loss, and local refinement as contributions for effective segmentation of long videos.It evaluates the method on public benchmarks and emphasizes practical long-video applications.

2 Related Work

Related VOS methods learn object appearance implicitly through network weights or explicitly through an embedding space and similarity-based pixel classification. The paper positions adaptive feature-bank management and uncertainty refinement against these approaches.

  • 2 Related Work: Implicit VOS methods learn object features through network weights, using detection-based or propagation-based segmentation pipelines.They often require online learning to adapt the weights to new objects during testing.
  • 2 Related Work: Explicit, matching-based methods construct an embedding space for object appearance and classify current-frame pixels by feature similarity.Their central design issue is how to build and maintain the embedding space.
  • 2 Related Work: Existing explicit methods use fixed memory policies such as first-frame, first-and-latest-frame, sliding-window, or uniform sampling strategies.For long videos, these static strategies can miss key frames or encounter out-of-memory crashes.
  • 2 Related Work: Recent image-segmentation methods refine uncertain regions, while this work adapts uncertainty-based boundary refinement to video segmentation.The related approaches motivate fine-grained local refinement for ambiguous mask regions.

3 Approach

The framework combines matching-based segmentation with an adaptive feature bank and uncertain-region refinement. It updates stored object features while refining ambiguous areas to support segmentation in long videos.

  • Matching-based segmentation: The matching-based pipeline encodes reference frames and object masks, matches query features against object-specific banks, and decodes the retrieved results into masks.Each object has an independent feature bank, while query and reference features are embedded into key and value spaces for matching.
  • Adaptive feature bank: About 90% of new features satisfy the merging operation, so only the remaining 10% need to be added to the bank each time.This reduces feature-bank growth while retaining distinct features for object representation.
  • Adaptive feature bank: The bank removes obsolete features with the least frequently used index when its size approaches a predefined budget.The LFU index uses feature-use counts and residence time, and the removal procedure is designed to keep the bank suitable for videos of arbitrary length.
  • Uncertain-region refinement: Uncertain-region refinement estimates ambiguity with a confidence loss and applies local fine-grained refinement to ambiguous mask regions.The method targets boundary uncertainty arising from low-resolution decoding and improves segmentation quality when trained with the confidence loss.

4 Training details

AFB-URR is pretrained on simulated videos generated from static images, then trained separately on each benchmark’s training videos. Benchmark training uses six-frame samples with resizing, cropping, augmentation, and limited object selection.

  • The model is pretrained on simulation videos generated from static image datasets, then further trained on each benchmark’s training videos.
  • Synthetic pretraining clips contain one first frame and five subsequent frames generated from the same image through data augmentation.
  • Benchmark training randomly selects six frames per video, resizes and crops them to 400 × 400 pixels, and samples at most three objects.

5 Experiments

The experiments evaluate AFB-URR on standard and long-video benchmarks, showing strong accuracy, runtime, generalization, and long-video performance. Ablations support the contributions of adaptive memory management and uncertain-region refinement.

  • Evaluation setup: The DAVIS17 evaluation uses region accuracy J, boundary accuracy F, mean score M, recall R, and decay D to assess segmentation quality.J is mask IoU, F measures boundary accuracy, R uses threshold τ = 0.5, and D measures change over time.
  • State-of-the-art comparison: AFB-URR achieves 74.6 J&F on DAVIS17 without online fine-tuning, compared with STM’s 71.6, while running at 4.0fps versus 3.4fps.With memory usage limited under 20%, it reaches 5.7fps and 71.7 J&F.
  • State-of-the-art comparison: 79.6 overall score on YouTube-VOS is the best reported result, while unseen-object performance reaches 74.1 J and 82.6 F.STM performs somewhat better on already-seen objects, under different memory budgets and hardware.
  • Long-time video comparison: 83.3 J&F on the Long-time Video dataset is the best score, exceeding STM’s 79.3 after videos extend beyond the fixed memory capacity.RVOS and A-GAME failed to recognize the object after 1K frames, while STM’s fixed 50-frame storage increases key-frame intervals for longer videos.
  • Ablation study: AFB outperforms first-frame, latest-frame, and manually combined reference-frame strategies by organizing key information from previous frames more effectively.The full AFB+URR framework has the best performance in the memory-management ablation.
  • Ablation study: Uncertain-region refinement improves performance in ambiguous boundary regions through uncertainty evaluation and local refinement.The ablation disables the confidence loss and/or local refinement to assess URR’s contribution.

6 Conclusion

The paper presents AFB-URR for semi-supervised video object segmentation, combining adaptive feature organization with uncertain-region refinement. It reports improved performance over state-of-the-art methods on two large-scale benchmarks.

  • AFB-URR combines an adaptive feature bank for organizing key segmentation features with uncertain-region refinement for ambiguous regions.Training minimizes segmentation cross-entropy together with a confidence loss.

Broader Impact

The framework is intended to support long-video processing tasks beyond VOS, including autonomous driving, robot interaction, and surveillance monitoring. It also illustrates flood detection and monitoring as a potential societal application.

  • The adaptive feature bank and matching framework can be modified for video processing tasks involving long videos and changing appearances.
  • Potential application areas include autonomous driving, robot interaction, and video surveillance monitoring.
  • Flood monitoring could use video segmentation to locate flooding and estimate water levels from surveillance-camera footage.
Loading 2010.07958v1…