PaperlayerView exact revision on arXiv

Source-linked AI summary

Contextual Diversity for Active Learning

Sharat Agarwal, Himanshu Arora, Saket Anand, Chetan Arora

arXiv:2008.05723v1cs.CV

TL;DR

Large annotation requirements and existing uncertainty or visual-diversity measures that miss spatial context motivate contextual-diversity active learning. The paper introduces an information-theoretic measure from CNN softmax predictions and applies it through core-set and reinforcement-learning selection. Across semantic segmentation, object detection, and image classification, the approaches report state-of-the-art results, with RL-based selection improving over state of the art by 2.7, 2.1, and 2.3 units, respectively.

  • Problem

    Existing active-learning measures based on prediction uncertainty or global visual diversity do not adequately capture spatial and semantic context used by CNNs.

  • Method

    Contextual diversity computes an information-theoretic distance over mixtures of softmax distributions from pseudo-labeled regions and supports core-set and reinforcement-learning selection.

  • Results

    Contextual diversity achieves state-of-the-art active-learning performance across semantic segmentation, object detection, and image classification; CDAL-RL improves over state of the art by 2.7, 2.1, and 2.3 units, respectively.

  • Takeaways & Limitations

    Contextual diversity captures predictive uncertainty together with class-specific confusion arising from spatial and semantic context.

  • Takeaways & Limitations

    The method assumes a sufficiently large unlabeled pool with a non-empty region set for each class, and deeper theoretical interpretations remain future work.

Abstract

from arXiv · show

Requirement of large annotated datasets restrict the use of deep convolutional neural networks (CNNs) for many practical applications. The problem can be mitigated by using active learning (AL) techniques which, under a given annotation budget, allow to select a subset of data that yields maximum accuracy upon fine tuning. State of the art AL approaches typically rely on measures of visual diversity or prediction uncertainty, which are unable to effectively capture the variations in spatial context. On the other hand, modern CNN architectures make heavy use of spatial context for achieving highly accurate predictions. Since the context is difficult to evaluate in the absence of ground-truth labels, we introduce the notion of contextual diversity that captures the confusion associated with spatially co-occurring classes. Contextual Diversity (CD) hinges on a crucial observation that the probability vector predicted by a CNN for a region of interest typically contains information from a larger receptive field. Exploiting this observation, we use the proposed CD measure within two AL frameworks: (1) a core-set based strategy and (2) a reinforcement learning based policy, for active frame selection. Our extensive empirical evaluation establish state of the art results for active learning on benchmark datasets of Semantic Segmentation, Object Detection and Image Classification. Our ablation studies show clear advantages of using contextual diversity for active learning. The source code and additional results are available at https://github.com/sharat29ag/CDAL.

1 Introduction

Large annotation costs motivate active learning, but existing uncertainty and visual-diversity measures miss spatial context used by CNNs. The paper introduces contextual diversity and applies it to core-set and reinforcement-learning selection, reporting improvements across three visual tasks.

  • Motivation: Large labeled datasets make CNN deployment costly across tasks and domains with different annotation requirements.Annotation effort varies substantially across image classification, object detection, semantic segmentation, and specialized imaging domains.
  • Motivation: Active learning selects informative samples under a labeling budget, and models trained on selected subsets can approach full-data performance.The approach leverages the model’s current knowledge rather than choosing samples randomly.
  • Research gap: Existing uncertainty measures can select correlated samples, while global visual-diversity features omit object locations and relative spatial relationships.Entropy also does not identify which classes cause uncertainty.
  • Proposed idea: Contextual diversity targets uncertainty arising from diverse spatial and semantic contexts because CNN receptive fields incorporate surrounding objects.The objective is to select frames containing objects in varied contexts and backgrounds.
  • Results: 2.7, 2.1, and 2.3 units improve the respective tasks over state of the art when contextual diversity is used as an RL reward.The paper also reports that contextual diversity complements visual diversity in ablation experiments.

2 Related Work

Prior active-learning methods use committees, diversity, uncertainty, or learned representations, but often face computational, diversity, or task-alignment limitations. The proposed framework combines contextual-diversity selection with core-set and reinforcement-learning modules.

  • Existing approaches: Query-by-committee methods use consensus among several models but are generally too computationally expensive for deep networks and large datasets.This limits their practicality in the setting motivating the paper.
  • Existing approaches: Diversity-based methods seek representative subsets through clustering, matrix partitioning, or linkage-based similarity, while uncertainty methods exploit ambiguity in model predictions.Some methods combine both diversity and uncertainty cues.
  • Core-set methods: Core-set active learning selects points in CNN feature space with Euclidean distances and offers guarantees tied to the covering radius.The paper replaces this representation and distance with contextual diversity in probability space.
  • Uncertainty methods: Learning-Loss sampling ranks examples by predicted loss, but high-loss samples may be outliers or label noise and may not yield the greatest performance gain.The cited limitation concerns whether predicted loss identifies the most useful samples.
  • Representation-based methods: VAAL selects samples using a discriminator in a learned latent space, without an obvious mechanism for choosing diverse samples or ensuring alignment with the task model.VAAL nevertheless reports state-of-the-art performance for image classification and semantic segmentation in the cited empirical analysis.
  • Proposed framework: The proposed frame-selection architecture supports unsupervised selection through either CDAL-CS or CDAL-RL, using a pretrained task model before annotation and fine-tuning.The same selection setup is presented as generalizable across visual tasks.

3 Active Frame Selection

The paper defines contextual diversity from class-specific confusion in CNN predictions and uses it to select informative, contextually diverse frames. It implements this measure in core-set and reinforcement-learning strategies, with additional visual and semantic representation rewards in the RL framework.

  • Contextual Diversity: Entropy-weighted class-specific confusion captures uncertainty arising from spatial and semantic context in CNN receptive fields.Higher entropy indicates probability mass spread across classes, reflecting confusion beyond the pseudo-labeled class.
  • Contextual Diversity: Contextual diversity measures confusion between classes using mixtures of softmax predictions over pseudo-labeled regions.Regions may be pixels, bounding boxes, or entire images, allowing the measure to support semantic segmentation, object detection, and image classification.
  • Contextual Diversity: Pairwise contextual diversity uses symmetric KL-divergence to quantify differences in class-specific confusion between two images.An indicator includes a class only when both images contain at least one region pseudo-labeled with that class, providing a somewhat reliable comparison.
  • Frame Selection Strategies: CDAL-CS replaces Euclidean core-set distances with pairwise contextual diversity in the K-Center-Greedy selection algorithm.The resulting representation is the per-image mixture distribution rather than a feature embedding.
  • Frame Selection Strategies: CDAL-RL trains a Bi-LSTM policy with REINFORCE, using contextual diversity as a reward alongside visual and semantic representation components.Visual representation spreads selections across feature space, while semantic representation balances classes and is used only for semantic segmentation.
  • Network Architecture and Training: The contextual diversity measure is task-network agnostic and is computed from predicted softmax probabilities.For CDAL-RL, the policy input is a vectorized nC × nC contextual representation extracted using the task network.

4 Results and Comparison

The evaluation compares contextual-diversity active learning with established approaches across semantic segmentation, object detection, and image classification. CDAL variants achieve strong performance while reducing the labeled data needed for target results.

  • Experimental Setup: The study evaluates active learning on semantic segmentation, object detection, and image classification using established experimental protocols and task-specific metrics.It uses Cityscapes and BDD100K with mIoU, PASCAL VOC with mAP, and CIFAR-10/CIFAR-100 with accuracy.
  • Semantic Segmentation: CDAL variants outperform competing approaches on Cityscapes and BDD100K, reaching 57.2 and 42.8 mIoU while reducing labeling by 300 and 800 frames, respectively.The comparison uses the DRN backbone and evaluates budgets from 10% to 40% of the unlabeled pool.
  • Object Detection: CDAL-RL achieves 73.3 mAP with 8k labeled samples, whereas learning loss reaches the same value with 10k, reducing labeling by 2k samples.CDAL performs nearly as well as core-set early and surpasses competing approaches in later active-learning cycles.
  • Image Classification: CDAL convincingly outperforms VAAL on CIFAR-10 and CIFAR-100, achieving approximately 81% CIFAR-10 accuracy with 5,000 fewer samples and beating 47.95% CIFAR-100 accuracy with 2,500 fewer samples.The paper attributes this scaling behavior to contextual diversity’s accumulated KL-divergence computation.
  • Contextual Diversity: In image classification, contextual diversity aggregates confusion among classes and selects batches diverse in both classes and their confusion.The entire image serves as a region, so a single-component probability mixture still captures spatial-context confusion.

5 Analysis and Ablation Experiments

Ablation experiments on Cityscapes examine reward components, policy-training frequency, feature-space representativeness, and the number of classes contributing to contextual diversity.

  • Policy Training Analysis: The experiments also test the effect of learning the Bi-LSTM policy only once during the first active-learning iteration.The corresponding Cityscapes comparison uses 10% randomly selected frames initially.
  • Experimental Setting: The ablations use Cityscapes semantic segmentation with DRN because this task has especially high annotation time per image.All experiments follow the settings used for the main semantic-segmentation evaluation.
  • Reward Component Ablation: Contextual diversity alone outperforms VAAL, and adding visual and semantic representation terms further improves the reward-based approach.The compared rewards are contextual diversity alone, contextual diversity plus visual representation, and the full three-component reward.
  • Feature-Space Analysis: CDAL selects points more uniformly across the feature space than core-set, indicating stronger representativeness under the reported comparison.The analysis links core-set’s behavior to limitations of Euclidean distance in the DRN feature space.
  • Class-wise Contextual Diversity: Adding contextual-diversity reward for Person and Vegetation substantially improves their IoU scores, while Person-only reward yields a substantial Person gain and marginal overall improvement.The analysis computes contextual diversity over an increasing number of classes.

6 Conclusion

The paper introduces contextual diversity as an information-theoretic measure for active frame selection, using it with core-set and reinforcement-learning frameworks. The measure captures predictive uncertainty and class-specific confusion through pseudo-labeled data-point distributions.

  • Contextual diversity serves as a distance measure for core-set selection and a reward function for reinforcement-learning-based active frame selection.These uses are evaluated across three visual recognition tasks.
  • The information-theoretic measure is computed over mixtures of softmax distributions from pseudo-labeled data points.This design captures both model predictive uncertainty and class-specific confusion.
  • The paper reports promising empirical results while leaving deeper theoretical interpretations of contextual diversity for future investigation.

7 Region Level Selection

The region-level comparison evaluates CDAL-RL against CEREALS and RBAL on Cityscapes. The reported table comparison states that CDAL-RL outperforms both methods by a significant margin.

  • CDAL-RL outperforms CEREALS and RBAL on Cityscapes in comparisons with region-based active learning approaches.The comparison uses the respective architectures to compute contextual diversity for frame selection.

8 Qualitative Results for CDAL

The qualitative results examine how selecting different class sets for the contextual-diversity reward changes the frames chosen by CDAL-RL. Including more classes reduces corresponding class confusion and changes the represented spatial neighborhoods.

  • The ablation compares CDAL-RL selections using {Sidewalk}, {Sidewalk, Fence}, and {Sidewalk, Fence, Vegetation} in the reward.Top-three frames from three independent runs are shown for each class set.
  • Adding classes to contextual-diversity computation reduces corresponding class confusion, reflected by lower entropy and peakier mixture distributions.
  • The selected spatial neighborhoods around Sidewalk include different classes such as Car, Motorcycle, and Person across the class-set conditions.

9 Ablation on Image Classification: CIFAR100

The CIFAR-100 ablations test CDAL-RL under biased initial pools, noisy labels, varying budgets, and a different network architecture. Across these settings, the reported results show CDAL-RL remaining better than existing or competitive approaches.

  • Biased Initial Pool: CDAL-RL performs better than existing techniques when the initial labeled pool excludes data from 10 or 20 classes.
  • Noisy Oracle: CDAL-RL is substantially more robust to 10%, 20%, and 30% noisy labels than other approaches.The selected labels are replaced with random classes from the same super-class; the authors associate robustness with pairwise KL-divergence-based frame selection.
  • Varying Budget: Varying the budget step size to 10% does not substantially affect CDAL-RL performance, which remains better than VAAL and competitive approaches.
  • Change in network Architecture: On CIFAR-100 with ResNet18, CDAL-RL outperforms all existing baselines by a substantial margin.

10 Ablation: Sensitivity analysis of α

CDAL-RL’s reward weighting is relatively robust: α = 0.75 gives the highest Cityscapes mIoU, while tested alternatives remain above competing methods. Replacing Shannon’s entropy weighting with mixture weights reduces performance.

  • α weighting sensitivity: 55.5% mIoU is the lowest tested CDAL-RL result, yet it exceeds VAAL (∼54%) and CDAL-CS (∼54.9%).The authors report this under the tested α values 0.25, 0.5, and 0.75.
  • α weighting sensitivity: α = 0.75 achieves the highest mIoU among the tested reward-weighting settings.The experiments compare α values of 0.25, 0.5, and 0.75 for selecting the next 5% of samples.
  • Alternative weighting: 56.3% mIoU falls to 55.2% when mixture weights replace Shannon’s entropy in the Cityscapes experiment.The authors attribute the deterioration to Shannon’s entropy better capturing prediction uncertainty.

11 Algorithm

The algorithms select active samples using contextual-diversity-based representations and either core-set distances or a reinforcement-learning policy. The accompanying experiments vary class sets, initialization, oracle noise, and selection budgets, with CIFAR-100 performance compared using ResNet-18.

  • CDAL-CS: CDAL-CS initializes a selected pool randomly, then iteratively chooses the unlabeled point farthest from the nearest selected centroid.Distances are computed from the feature matrix using Eq. (2) until the budget is reached.
  • CDAL-RL: CDAL-RL predicts probabilities for each data point, samples using the highest probabilities, computes three rewards, and updates its policy with REINFORCE and SGD.The procedure repeats across epochs to return trained reinforcement-learning parameters.
  • Ablation experiments: Qualitative experiments compare contextual diversity computed from Sidewalk, Sidewalk and Fence, or Sidewalk, Fence and Vegetation.The figure pairs selected frames with ground truth and prediction overlays.
  • Ablation experiments: CDAL-RL ablations vary biased initial-pool sizes, noisy-oracle conditions, and selection budgets on CIFAR-100.The figure specifies m=10, m=20, a noisy oracle, and budget steps of 10%.
  • Evaluation: ResNet-18 is used for the CIFAR-100 performance comparison of CDAL-RL.
Loading 2008.05723v1…