FCOS: Fully Convolutional One-Stage Object Detection (1904.01355v5)

Abstract: We propose a fully convolutional one-stage object detector (FCOS) to solve object detection in a per-pixel prediction fashion, analogue to semantic segmentation. Almost all state-of-the-art object detectors such as RetinaNet, SSD, YOLOv3, and Faster R-CNN rely on pre-defined anchor boxes. In contrast, our proposed detector FCOS is anchor box free, as well as proposal free. By eliminating the predefined set of anchor boxes, FCOS completely avoids the complicated computation related to anchor boxes such as calculating overlapping during training. More importantly, we also avoid all hyper-parameters related to anchor boxes, which are often very sensitive to the final detection performance. With the only post-processing non-maximum suppression (NMS), FCOS with ResNeXt-64x4d-101 achieves 44.7% in AP with single-model and single-scale testing, surpassing previous one-stage detectors with the advantage of being much simpler. For the first time, we demonstrate a much simpler and flexible detection framework achieving improved detection accuracy. We hope that the proposed FCOS framework can serve as a simple and strong alternative for many other instance-level tasks. Code is available at:Code is available at: https://tinyurl.com/FCOSv1

• The paper introduces an anchor-free framework that removes the need for anchor boxes, reducing hyper-parameter tuning and simplifying object detection.
• The novel center-ness branch down-weights low-quality predictions, effectively suppressing false positives and boosting overall detection accuracy.
• Multi-level prediction with FPN enables effective handling of various object sizes, achieving competitive AP and outperforming traditional one-stage detectors.

FCOS: Fully Convolutional One-Stage Object Detection

The paper "FCOS: Fully Convolutional One-Stage Object Detection" presents a novel approach to object detection, diverging from the prevalent use of anchor boxes in state-of-the-art detectors like RetinaNet, SSD, YOLOv3, and Faster R-CNN. The proposed FCOS (Fully Convolutional One-Stage Object Detector) aims to simplify the object detection process by adopting an anchor-free and proposal-free framework, reminiscent of dense prediction tasks such as semantic segmentation.

Key Contributions

The primary contributions of the FCOS framework are:

1. Anchor-Free Detection: FCOS eliminates the dependence on anchor boxes, thus avoiding the associated hyper-parameters and the complex computation of overlaps during training. This simplification leads to improved performance and reduced design complexity.
2. Center-Ness Branch: The introduction of a novel "center-ness" branch helps suppress low-quality bounding box predictions by down-weighting the scores of boxes that are far from the object center. This addition significantly improves detection accuracy.
3. Multi-Level Prediction with FPN: By leveraging feature pyramid networks (FPN), FCOS handles objects of varying sizes more effectively, alleviating issues related to overlapping bounding boxes and improving best possible recall (BPR).
4. Competitive Performance: FCOS demonstrates state-of-the-art performance among one-stage detectors, surpassing traditional anchor-based detectors in both accuracy and simplicity.

Detailed Overview

Anchor-Free Approach

Traditional object detectors rely on pre-defined anchor boxes to generate possible bounding boxes for objects within an image. This reliance introduces several challenges, including sensitivity to hyper-parameters related to anchor box sizes, aspect ratios, and the number of anchor boxes, which can significantly impact detection performance.

FCOS removes the need for anchor boxes by treating each pixel as a potential object center and predicting bounding boxes directly. Specifically, at each pixel location on the feature map, the network predicts a 4D vector representing the bounding box coordinates relative to the pixel and a classification score for object categories.

Center-Ness Branch

The center-ness branch is a crucial addition to FCOS. It predicts a normalized distance from each pixel to the center of the object, which is then used to down-weight the scores of bounding boxes that are likely to be of low quality. This approach helps in effectively filtering out false positives, particularly for overlapping objects, thus enhancing the overall detection accuracy.

Multi-Level Prediction

To address the issues of recall and overlapped bounding boxes, FCOS employs a multi-level prediction strategy using FPN. Different feature pyramid levels handle objects of different sizes, reducing the ambiguity caused by overlapping bounding boxes. This approach leads to a high BPR, comparable to or better than traditional anchor-based methods.

Performance and Comparisons

FCOS achieves remarkable performance on the MS-COCO benchmark. With a ResNeXt-64x4d-101 backbone, FCOS attains an AP of 44.7% under single-model and single-scale testing, which outperforms many state-of-the-art one-stage detectors, including RetinaNet and CornerNet. The proposed framework also demonstrates competitive recall rates, with substantial improvements over traditional anchor-based methods.

Practical and Theoretical Implications

The simplicity and effectiveness of the FCOS framework have several implications:

• Reduced Complexity: By eliminating the need for anchor boxes, FCOS significantly reduces the design complexity and the number of hyper-parameters, simplifying the training process and reducing the chances of overfitting due to excessive parameter tuning.
• Unified Framework: FCOS unifies object detection with other dense prediction tasks, promoting greater reuse of ideas and methods across different visual recognition problems.
• Versatility: The anchor-free approach of FCOS can be readily extended to other instance-level tasks, such as instance segmentation and key-point detection, with minimal modifications.

Future Developments

The promising results of FCOS suggest several future research directions:

• Exploring Variants and Extensions: Further exploration into different architectural variants and extending the anchor-free methodology to a broader range of vision tasks can lead to more robust and versatile detectors.
• Optimization and Fine-Tuning: While FCOS already achieves competitive performance, fine-tuning the hyper-parameters specifically for the FCOS framework could yield even higher accuracy.
• Integration with Advanced Backbones: Evaluating FCOS with different backbone networks, such as EfficientNets or Transformers, may provide insights into achieving better trade-offs between accuracy and computational efficiency.

In conclusion, FCOS presents a compelling case for revisiting the necessity and utility of anchor boxes in object detection. By demonstrating that an anchor-free approach can achieve superior performance with reduced complexity, FCOS sets a new benchmark for the development of future object detection systems.