Source-linked AI summary
Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection
Xiang Li, Wenhai Wang, Xiaolin Hu, Jun Li, Jinhui Tang, Jian Yang
TL;DR
Dense object detectors need reliable localization-quality scores for ranking boxes during NMS, while prior methods commonly derive these scores from vanilla convolutional features. This paper uses statistics of learned bounding-box distributions through a lightweight Distribution-Guided Quality Predictor, producing GFLV2. On COCO test-dev, GFLV2 reaches 46.2 AP versus 43.6 AP for ATSS, a 2.6 AP gain without sacrificing training or inference efficiency.
Problem
Dense object detection needs reliable Localization Quality Estimation to rank boxes accurately during Non-Maximum Suppression, while existing methods commonly use vanilla convolutional features.
Method
GFLV2 uses statistics of learned bounding-box distributions in a lightweight Distribution-Guided Quality Predictor to produce localization-quality scores within the GFLV1 framework.
Results
46.2 AP on COCO test-dev exceeds the 43.6 AP ATSS baseline by 2.6 AP with ResNet-101, without sacrificing training or inference efficiency.
Takeaways & Limitations
Bounding-box distribution statistics provide a lightweight and efficient basis for reliable LQE in dense object detection.
Abstract
from arXiv · showhide
Localization Quality Estimation (LQE) is crucial and popular in the recent advancement of dense object detectors since it can provide accurate ranking scores that benefit the Non-Maximum Suppression processing and improve detection performance. As a common practice, most existing methods predict LQE scores through vanilla convolutional features shared with object classification or bounding box regression. In this paper, we explore a completely novel and different perspective to perform LQE -- based on the learned distributions of the four parameters of the bounding box. The bounding box distributions are inspired and introduced as "General Distribution" in GFLV1, which describes the uncertainty of the predicted bounding boxes well. Such a property makes the distribution statistics of a bounding box highly correlated to its real localization quality. Specifically, a bounding box distribution with a sharp peak usually corresponds to high localization quality, and vice versa. By leveraging the close correlation between distribution statistics and the real localization quality, we develop a considerably lightweight Distribution-Guided Quality Predictor (DGQP) for reliable LQE based on GFLV1, thus producing GFLV2. To our best knowledge, it is the first attempt in object detection to use a highly relevant, statistical representation to facilitate LQE. Extensive experiments demonstrate the effectiveness of our method. Notably, GFLV2 (ResNet-101) achieves 46.2 AP at 14.6 FPS, surpassing the previous state-of-the-art ATSS baseline (43.6 AP at 14.6 FPS) by absolute 2.6 AP on COCO {\tt test-dev}, without sacrificing the efficiency both in training and inference. Code will be available at https://github.com/implus/GFocalV2.
1. Introduction
Localization Quality Estimation improves dense detection by ranking higher-quality boxes above lower-quality ones during NMS. GFLV2 uses bounding-box distribution statistics to predict reliable quality scores with a lightweight predictor, achieving strong COCO results without sacrificing efficiency.
- Localization Quality Estimation ranks higher-quality bounding boxes above lower-quality ones, reducing mistaken suppression during Non-Maximum Suppression.
- Previous LQE methods generally estimate quality from vanilla convolutional features associated with points, borders, or regions.
- Bounding-box distribution statistics are strongly correlated with real localization quality, motivating their use instead of vanilla convolutional features.
- DGQP uses a lightweight sub-network with only dozens of hidden units, such as 64, to produce reliable LQE scores with negligible additional computation.
- 46.2 AP on COCO test-dev gives GFLV2 a 2.6 AP gain over the 43.6 AP ATSS baseline with ResNet-101, under the same training schedule and without sacrificing training or inference efficiency.
- The work presents bridging bounding-box distribution statistics and LQE in an end-to-end dense detection framework as a first attempt.
- GFLV2 is reported as lightweight, plug-compatible with most dense detectors, and providing approximately 2 AP gains without loss of training or inference speed.
- GFLV2 achieves 53.3 AP with Res2Net-101-DCN under multi-scale testing on COCO.
2. Related Work
Prior LQE methods use confidence, center distance, IoU, or other convolutional features to rank detections. GFLV2 instead uses statistics of learned bounding-box distributions, which are reported to combine high accuracy with efficiency.
- Early detectors used classification confidence for LQE, while later methods added localization features or alternative quality formats such as Centerness and boundary-bucketing confidence.
- IoU became a widely studied LQE measurement, with several methods using separate branches and GFLV1 and VFNet using a Classification-IoU joint representation.
- GFLV2 differs from prior approaches by predicting LQE from directly correlated statistics of bounding-box distributions rather than convolutional features.
- The statistics of bounding-box distributions achieve high accuracy and efficiency simultaneously compared with convolutional feature inputs.
3. Method
GFLV2 performs localization quality estimation from statistics of learned bounding-box distributions rather than vanilla convolutional features. It combines this distribution-guided IoU estimate with classification through a joint representation used for training and NMS inference.
- General Distribution: GFLV1 represents each bounding-box edge with a discrete General Distribution whose expected value provides the regression estimate.The distribution is defined over an evenly discretized predefined output range, with probabilities summing to one.
- Distribution-Guided Quality Estimation: The General Distribution reflects prediction quality, motivating distribution statistics as inputs for localization quality estimation.Its flatness is related to final bounding-box quality: sharper distributions correspond to higher quality, while flatter distributions correspond to lower quality.
- Classification-IoU Joint Representation: GFLV2 decomposes the Classification-IoU representation as J = C × I, combining classification representation C with a scalar IoU representation I.The same joint representation is supervised by Quality Focal Loss during training and used directly as the NMS score during inference.
- Distribution-Guided Quality Predictor: DGQP extracts each edge distribution’s Top-k values and mean, concatenates them into F ∈ R4(k+1), and predicts I with a two-layer ReLU-Sigmoid network.A typical setting uses k = 4 and hidden dimension p = 64.
- Distribution Statistics: Top-k and mean statistics reflect distribution flatness while remaining insensitive to relative offsets, making the feature robust to object scales.Because each distribution sums to a fixed value, larger selected values indicate sharper distributions and smaller values indicate flatter ones.
4. Experiment
Experiments show that distribution statistics provide effective, lightweight guidance for LQE, improving accuracy, correlation, training behavior, and NMS ranking without sacrificing efficiency.
- Ablation Study: Top-4 distribution values combined with their mean perform best as DGQP input, with k = 4 and p = 64 achieving optimal accuracy.These settings are used as the standard configuration in subsequent experiments.
- Ablation Study: Distribution statistics achieve the best overall AP and fastest inference compared with evaluated convolutional input features.The comparison covers point, region, border, regular-point, and deformable-point features.
- Ablation Study: The Decomposed Form consistently outperforms Composed Forms across dimension settings in both accuracy and running speed.The Decomposed Form represents the joint score as J = C × I, while Composed Forms use enriched inputs to fully connected layers.
- Analysis: DGQP improves estimated-IoU correlation by +0.26 over GFLV1 and produces an absolute 0.9 AP gain.The comparison uses Pearson Correlation Coefficients between estimated and ground-truth IoUs on COCO minival positive samples.
- Analysis: GFLV2 maintains approximately 41 AP while preserving training and inference efficiency, unlike several compared detectors with reported time overheads.PAA, RepPointsV2, and BorderDet add 52%, 65%, and 22% training overheads, respectively.
5. Conclusion
The paper proposes reliable LQE guided by bounding-box distribution statistics and develops GFLV2 as a lightweight dense detector. COCO experiments validate its effectiveness, compatibility, and efficiency.
- 5. Conclusion: GFLV2 learns reliable localization quality estimation through statistics of bounding-box distributions.The approach is presented as a new perspective for dense object detection.
- 5. Conclusion: Experiments and analyses on COCO validate GFLV2's effectiveness, compatibility, and efficiency.The paper presents GFLV2 as a simple yet effective baseline for the community.