CFA: Constraint-based Finetuning Approach for Generalized Few-Shot Object Detection

Guirguis, Karim; Hendawy, Ahmed; Eskandar, George; Abdelsamad, Mohamed; Matthias, Kayser,; Beyerer, Jürgen

doi:10.1109/cvprw56347.2022.00449

Cited by 19 publications

(6 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…MPSR [24] was proposed to compensate for the lack of positive samples using data augmentation. DeFRCN [25] and CFA [26] improved network performance from the perspective of loss gradients. Compared with few-shot object detection on natural semantic images, object detection in RSIs is subject to additional challenges in more diverse scale and orientation variations.…”

Section: B Few-shot Object Detectionmentioning

confidence: 99%

Few-Shot Object Detection on Remote Sensing Images

Deng

Fang

2022

IEEE Trans. Geosci. Remote Sensing

View full text Add to dashboard Cite

Object detection in remote sensing images relies on a large amount of labeled data for training. The growing new categories and class imbalance render exhaustive annotation nonscalable. Few-shot object detection (FSOD) tackles this issue by meta-learning on seen base classes and then fine-tuning on novel classes with few labeled samples. However, the object's scale and orientation variations are particularly large in remote sensing images, thus posing challenges to existing few-shot object detection methods. To tackle these challenges, we first propose to integrate a feature pyramid network and use prototype features to highlight query features to improve upon existing FSOD methods. We refer to the modified FSOD as a Strong Baseline which is demonstrated to perform significantly better than the original baselines. To improve the robustness of orientation variation, we further propose a transformation-invariant network (TINet) to allow the network to be invariant to geometric transformations. Extensive experiments on three widely used remote sensing object detection datasets, i.e., NWPU VHR-10.v2, DIOR, and HRRSD demonstrated the effectiveness of the proposed method. Finally, we reproduced multiple FSOD methods for remote sensing images to create an extensive benchmark for follow-up works.

show abstract

Section: B Few-shot Object Detectionmentioning

confidence: 99%

Few-Shot Object Detection on Remote Sensing Images

Deng

Fang

2022

IEEE Trans. Geosci. Remote Sensing

View full text Add to dashboard Cite

show abstract

“…Guirguis et al [123] (CFA) built on the continual learning approaches GEM [134] and A-GEM [135], which observed that catastrophic forgetting occurs when the angle between loss gradient vectors of previous tasks and the gradient update of the current task is obtuse. Therefore, CFA stores K shots of the base categories in episodic memory, analogous to A-GEM, in order to be able to compute gradients on D base .…”

Section: E Keep the Performance On Base Categoriesmentioning

confidence: 99%

“…Takeaway: To prevent catastrophic forgetting and keep the performance on base categories, the angle between gradients of novel and base categories must be considered [123].…”

Section: E Keep the Performance On Base Categoriesmentioning

confidence: 99%

“…Guirguis et al [123] (CFA) derived a new gradient update rule that takes the angle between gradients for samples of D base and samples of D novel into account in order to combat catastrophic forgetting for base categories C base during the finetuning phase, as already described in Section VII-E. While this gradient update rule primarily intends to preserve the performance on base categories C base , it also has a positive impact on the performance regarding novel categories C novel .…”

Section: F Modify the Training Objectivementioning

confidence: 99%

“…CFA [123] can be applied on top of DeFRCN to tackle catastrophic forgetting of base categories during fine-tuning, which occurs when angles between base category gradients and novel category gradients are obtuse. Therefore, CFA stores K shots of the base categories in episodic memory for computing gradients on D base during fine-tuning.…”

Section: Summary Of Best Performing Transfer Learning Approachesmentioning

confidence: 99%

See 2 more Smart Citations

Few-Shot Object Detection: A Comprehensive Survey

Köhler

Eisenbach

Groß

2024

IEEE Trans. Neural Netw. Learning Syst.

View full text Add to dashboard Cite

Humans are able to learn to recognize new objects even from a few examples. In contrast, training deep-learningbased object detectors requires huge amounts of annotated data. To avoid the need to acquire and annotate these huge amounts of data, few-shot object detection (FSOD) aims to learn from few object instances of new categories in the target domain.In this survey, we provide an overview of the state of the art in FSOD. We categorize approaches according to their training scheme and architectural layout. For each type of approach, we describe the general realization as well as concepts to improve the performance on novel categories. Whenever appropriate, we give short takeaways regarding these concepts in order to highlight the best ideas. Eventually, we introduce commonly used datasets and their evaluation protocols and analyze the reported benchmark results. As a result, we emphasize common challenges in evaluation and identify the most promising current trends in this emerging field of FSOD.

show abstract

Robust Feature Space Organization with Distillation for Few-Shot Object Detection

Shenoy,

Chellappa

2024

Lecture Notes in Computer Science

View full text Add to dashboard Cite

CFA: Constraint-based Finetuning Approach for Generalized Few-Shot Object Detection

Cited by 19 publications

References 27 publications

Few-Shot Object Detection on Remote Sensing Images

Few-Shot Object Detection on Remote Sensing Images

Few-Shot Object Detection: A Comprehensive Survey

Robust Feature Space Organization with Distillation for Few-Shot Object Detection

Contact Info

Product

Resources

About