Box-Level Active Detection (2303.13089v1)

Abstract: Active learning selects informative samples for annotation within budget, which has proven efficient recently on object detection. However, the widely used active detection benchmarks conduct image-level evaluation, which is unrealistic in human workload estimation and biased towards crowded images. Furthermore, existing methods still perform image-level annotation, but equally scoring all targets within the same image incurs waste of budget and redundant labels. Having revealed above problems and limitations, we introduce a box-level active detection framework that controls a box-based budget per cycle, prioritizes informative targets and avoids redundancy for fair comparison and efficient application. Under the proposed box-level setting, we devise a novel pipeline, namely Complementary Pseudo Active Strategy (ComPAS). It exploits both human annotations and the model intelligence in a complementary fashion: an efficient input-end committee queries labels for informative objects only; meantime well-learned targets are identified by the model and compensated with pseudo-labels. ComPAS consistently outperforms 10 competitors under 4 settings in a unified codebase. With supervision from labeled data only, it achieves 100% supervised performance of VOC0712 with merely 19% box annotations. On the COCO dataset, it yields up to 4.3% mAP improvement over the second-best method. ComPAS also supports training with the unlabeled pool, where it surpasses 90% COCO supervised performance with 85% label reduction. Our source code is publicly available at https://github.com/lyumengyao/blad.

• The paper presents a box-level annotation framework that replaces inefficient image-level labeling with a targeted and resource-efficient approach.
• The paper introduces the Complementary Pseudo Active Strategy (ComPAS) which combines human annotations and model-generated pseudo-labels to focus on informative bounding boxes.
• The paper demonstrates superior performance, achieving full supervised results on PASCAL VOC0712 with only 19% annotation effort and a 4.3% mAP boost on COCO.

Overview of "Box-Level Active Detection" Paper

The paper "Box-Level Active Detection" introduces an innovative framework for object detection through active learning, addressing limitations in existing methodologies. Traditional active learning practices largely focus on image-level evaluation, where entire images are labeled and used for model training. This is not only inefficient in terms of human annotator workload but also leads to biased results favoring dense images. To overcome these deficiencies, this paper proposes a box-level active detection framework.

Key Contributions

The paper's core contributions can be summarized as follows:

1. Box-Level Annotation Framework: The authors propose a shift from image-level to box-level annotation. This approach aligns more closely with actual human workload, as tasks are measured by bounding boxes rather than whole images. It allows for more efficient allocation of annotation resources by focusing on informative targets while minimizing redundancy.
2. Complementary Pseudo Active Strategy (ComPAS): The framework integrates a novel pipeline called ComPAS, which combines human annotations with model-generated pseudo-labels. This strategy involves two main components:
   • An input-end committee that prioritizes informative bounding boxes for human annotation by analyzing prediction disagreements among different views of the input data.
   • A mechanism to generate pseudo-labels for well-understood targets, thereby ensuring only the truly informative samples require human intervention.
3. Superior Performance: ComPAS outperforms existing methods across various benchmarks. For instance, on the PASCAL VOC0712 dataset, ComPAS reaches full supervised performance using only 19% of the box annotations, achieving notable gains in efficiency and label reduction. On the larger COCO dataset, it provides an improvement of up to 4.3% in mean Average Precision (mAP) over the best competing methods.
4. Unified Codebase and Benchmarking: The authors provide a comprehensive codebase, ensuring that comparisons between different active learning approaches are conducted fairly and transparently, using the same basic detector architecture and experimental settings.

Theoretical and Practical Implications

Theoretical Contributions: The paper's shift to a box-level evaluation model presents a refined view of how active learning budgets could be more practically managed in computational tasks. The incorporation of pseudo-labels within the ComPAS framework demonstrates an effective balance between traditionally manual and automatic procedures in machine learning workflows.

Practical Implications: From a practical standpoint, this research indicates substantial reductions in the costs associated with labeling datasets, which directly affects overall model training costs. By applying fewer but more precise annotations, detection models can maintain high levels of accuracy without the overhead of exhaustive data labeling.

Future Directions

The introduction of a box-level active detection framework opens several avenues for future research, including:

• Exploration of more sophisticated disagreement metrics for even better informativeness estimation.
• Enhancing pseudo-label robustness through advanced confidence estimation techniques to improve the complementing role of pseudo-labels.
• Application of the proposed framework to other tasks in computer vision where annotation costs are high, further expanding the practicality of active learning solutions in diverse contexts.

In conclusion, the "Box-Level Active Detection" paper presents a notable advancement in the development of active learning methodologies for object detection, emphasizing efficiency and practicality. The proposed ComPAS pipeline demonstrates substantial potential for enhancing model training practices by reducing annotation demands while maintaining or improving model performance.