Incremental Few-Shot Object Detection (2003.04668v2)

Abstract: Most existing object detection methods rely on the availability of abundant labelled training samples per class and offline model training in a batch mode. These requirements substantially limit their scalability to open-ended accommodation of novel classes with limited labelled training data. We present a study aiming to go beyond these limitations by considering the Incremental Few-Shot Detection (iFSD) problem setting, where new classes must be registered incrementally (without revisiting base classes) and with few examples. To this end we propose OpeN-ended Centre nEt (ONCE), a detector designed for incrementally learning to detect novel class objects with few examples. This is achieved by an elegant adaptation of the CentreNet detector to the few-shot learning scenario, and meta-learning a class-specific code generator model for registering novel classes. ONCE fully respects the incremental learning paradigm, with novel class registration requiring only a single forward pass of few-shot training samples, and no access to base classes -- thus making it suitable for deployment on embedded devices. Extensive experiments conducted on both the standard object detection and fashion landmark detection tasks show the feasibility of iFSD for the first time, opening an interesting and very important line of research.

Incremental Few-Shot Object Detection

The research paper "Incremental Few-Shot Object Detection" presents a novel problem setting in the field of object detection termed Incremental Few-Shot Detection (iFSD). This problem setting addresses the limitations of current object detection models that require extensive labeled datasets and offline training. The proposed iFSD framework tackles the challenge of incrementally learning to recognize novel object classes from limited labeled examples without accessing the base class data.

Methodology

The authors introduce the OpeN-ended Centre nEt (ONCE), a detector designed for the iFSD problem. The development of ONCE is based on the adaptation of CentreNet, a one-stage detector known for its efficiency in object detection by predicting the heatmap of object centers. ONCE further exploits meta-learning techniques by incorporating a class-specific code generator to facilitate the incremental learning of new classes. The primary concept is to meta-learn a mechanism where novel class registration only requires a single forward pass with a few-shot support set.

Results

The ONCE approach is empirically validated on standard benchmarks such as COCO and PASCAL VOC. The experiments reveal that ONCE is capable of maintaining performance on base classes while addressing the challenge of catastrophic forgetting, a common issue in incremental learning tasks. Notably, in scenarios like fashion landmark detection on the DeepFashion2 dataset, ONCE demonstrates substantial capability, even with minimal examples per class.

Implications and Future Directions

This work is a significant step towards adaptive object detection systems that can operate effectively in dynamic environments with constrained data availability, such as embedded devices in robotic applications. The distinction of using class-specific modeling via independent heatmaps and meta-learned code generation sets a foundational basis for future explorations in incrementally learnable and scalable detection systems.

The implications of this research pave the way for embedded AI systems to autonomously adapt to novel scenarios with minimal human input. The ONCE model, with its no-backward-interference property achievable through independent class-specific detection, might inspire architectures in other areas requiring incremental updates, such as segmentation or tracking.

Further research could explore integrating ONCE with other weakly supervised or zero-shot learning techniques to exploit broader forms of input data, thereby enhancing its ability to generalize across vastly differing object classes without explicit labeled data. Another potential development could be investigating the transfer of knowledge across different domains or tasks to further leverage the learned feature representations.

In conclusion, the ONCE model offers a methodological advancement in object detection, addressing the practical challenges of incremental learning with few samples, proving beneficial for real-world applications where resource constraints are a critical concern.