Reducing Label Noise in Anchor-Free Object Detection (2008.01167v2)

Abstract: Current anchor-free object detectors label all the features that spatially fall inside a predefined central region of a ground-truth box as positive. This approach causes label noise during training, since some of these positively labeled features may be on the background or an occluder object, or they are simply not discriminative features. In this paper, we propose a new labeling strategy aimed to reduce the label noise in anchor-free detectors. We sum-pool predictions stemming from individual features into a single prediction. This allows the model to reduce the contributions of non-discriminatory features during training. We develop a new one-stage, anchor-free object detector, PPDet, to employ this labeling strategy during training and a similar prediction pooling method during inference. On the COCO dataset, PPDet achieves the best performance among anchor-free top-down detectors and performs on-par with the other state-of-the-art methods. It also outperforms all major one-stage and two-stage methods in small object detection (${AP}_{S}$ $31.4$). Code is available at https://github.com/nerminsamet/ppdet

• The paper introduces a novel sum-pooling method to reduce label noise in anchor-free object detection.
• The methodology enhances accuracy by pooling predictions from the central region to down-weight contributions from occluded and background features.
• PPDet achieves competitive results, notably 31.4 AP_S on the COCO dataset, marking significant improvements in small object detection.

Insights on Reducing Label Noise in Anchor-Free Object Detection

The paper addresses a critical issue in anchor-free object detection methodologies, focusing on the challenge of label noise stemming from imprecise labeling during training. Traditional anchor-free detectors label all features within a central region of a ground-truth box as positive, potentially categorizing background or occluder features as object features, leading to label noise. This noise adversely affects the detector’s performance by skewing the model's understanding of distinguishing object features from non-object features.

Proposed Method and Model

The authors introduce a novel approach to mitigate label noise by developing a new labeling strategy within anchor-free object detection frameworks. The main innovation is the sum-pooling of predictions derived from individual features into a single prediction during both training and inference phases. This method allows models to naturally down-weight the contributions from non-discriminatory features such as those on occluded parts or background areas.

The new object detector, PPDet, is crafted to utilize this labeling strategy effectively. During training, predictions within the "positive area," a centrally focused region of a ground-truth box, are pooled to create a unified prediction, thereby reducing the noisy label effect. This is a significant departure from current methods where each feature contributes separately to the loss function. During inference, PPDet employs a similar prediction pooling technique, strengthening the object hypotheses across overlapping predictions.

Results and Implications

PPDet exhibits competitive performance, particularly excelling in the small object detection as evidenced by an AP$_S$ of 31.4 on the COCO dataset, outperforming many state-of-the-art methods. This is notable given that improving small object detection has been a perennial challenge in object detection tasks.

The implications of this research extend to improved accuracy in real-world applications where robustness to label noise is critical—such as autonomous driving and surveillance. By focusing on reducing label noise, the methodology not only improves detection accuracy but could potentially lead to a reduction in training data requirements as more accurate models can be trained with less pristine datasets.

Future Speculations

Building on these foundations, future research may focus on further refinement of the pooling strategy to dynamically adjust based on the context or characteristics of the training data. Exploring adaptive pooling mechanisms or incorporating additional semantic information could offer pathways for enhanced performance. Additionally, leveraging this labeling strategy in conjunction with emerging neural architectures, such as transformers, might yield further improvements in object detection precision and robustness.

Overall, this paper contributes to the ongoing effort to refine feature discrimination in object detection systems, spotlighting the substantial gains achievable through strategic revisitation and reengineering of training methodologies for anchor-free detectors.