TL;DR
This paper introduces BGHR, a novel method that significantly improves oriented object detection by addressing the challenge of limited rotated bounding box (RBox) annotations. BGHR achieves this by adaptively mining high-quality training samples from readily available horizontal bounding box (HBox) annotations, effectively bridging the gap between HBox- and RBox-supervised methods.
What Problem Does This Paper Address?
Oriented object detection is crucial for accurately identifying objects in complex scenes like aerial imagery, where objects can appear at various angles. However, obtaining precise RBox annotations is expensive and time-consuming. While methods like H2RBox have shown promise in learning from cheaper HBox annotations, they primarily focus on the orientation information gap. This paper highlights a critical, overlooked gap: the difference in training sample selection between HBox- and RBox-supervised approaches. HBox-supervised methods often include noisy background samples, hindering performance. BGHR aims to solve this by intelligently selecting better training samples from HBox annotations.
Methods
The core of BGHR is the Adaptive Fine-grained Sample Mining (AFSM) strategy. AFSM works by:
- Prediction RBox Assignment (PRA): This scheme assigns the "best-matching" predicted RBox to each ground truth (GT) HBox. To determine the best match, it uses the Kullback-Leibler Divergence (KLD) as a metric for localization quality, converting bounding boxes into Gaussian distributions. Two assignment strategies are proposed: KLD-Only (KO) and Predictions-Mixture (PM), which also incorporates classification confidence.
- Improved Self-Supervised Branch Loss (Lss*): This component addresses the issue of symmetric prediction boxes in the weakly-supervised branch, which can lead to inconsistencies.
BGHR employs a two-branch architecture: a weakly-supervised (WS) branch (based on FCOS) for detection and a self-supervised (SS) branch that learns from two augmented views of the input image to enforce consistency. AFSM and PRA are integrated into the training process to guide the selection of high-quality samples for both branches.
Datasets
The paper evaluates BGHR on two standard datasets for oriented object detection:
- DOTA-v1.0
- DIOR-R
Both datasets provide annotations in both HBox and RBox formats, allowing for direct comparison with existing methods.
Key Insights and Ideas
- Sample Selection is as Crucial as Orientation: The paper convincingly argues that the quality of training samples is a significant bottleneck in HBox-supervised oriented object detection, not just the orientation representation itself.
- KLD for Robust RBox Matching: Using KLD to measure the similarity between predicted RBoxes and GT HBoxes proves more robust than traditional IoU-based metrics, especially when scale differences are large.
- Confidence-Aware Sample Mining: Incorporating classification confidence (in the PM assignment scheme) alongside localization quality leads to better sample selection, suggesting that the model's certainty about an object's class is a valuable signal for mining.
- Bridging the Gap Effectively: BGHR demonstrates that by focusing on sample mining, it can significantly close the performance gap with fully supervised RBox-based methods, even achieving competitive results.
Key Results
- BGHR achieves state-of-the-art performance among HBox-supervised methods on both DOTA-v1.0 and DIOR-R.
- On DOTA-v1.0, BGHR outperforms the baseline H2RBox by 3.62% in AP50, and with multi-scale training, it reaches 77.65% AP50.
- Compared to the RBox-supervised FCOS, BGHR achieves a slight performance advantage (0.66% AP50 higher on DOTA-v1.0).
- The proposed AFSM strategy improves AP75 by 2.85% and Lss* improves it by 1.97% when added to the baseline.
- The method incurs no additional computational overhead or detection speed difference compared to the baseline.
Limitations and Future Work
- Fine-grained Center-ness: The authors acknowledge that their approach still uses GT HBox to compute the center-ness target. They suggest that fine-grained center-ness prediction could further improve performance and leave this as future work.
- Hyperparameter Tuning: While the paper explores hyperparameter tuning for the PRA scheme, optimal settings might still require careful adjustment for specific applications.
Summary updated