Rilod

Abstract

Object detection models shipped with camera-equipped edge devices cannot cover the objects of interest for every user. Therefore, the incremental learning capability is a critical feature for a robust and personalized object detection system that many applications would rely on. In this paper, we present an efficient yet practical system, RILOD, to incrementally train an existing object detection model such that it can detect new object classes without losing its capability to detect old classes. The key component of RILOD is a novel incremental learning algorithm that trains end-to-end for one-stage deep object detection models only using training data of new object classes. Specifically to avoid catastrophic forgetting, the algorithm distills three types of knowledge from the old model to mimic the old model's behavior on object classification, bounding box regression and feature extraction. In addition, since the training data for the new classes may not be available, a real-time dataset construction pipeline is designed to collect training images on-the-fly and automatically label the images with both category and bounding box annotations. We have implemented RILOD under both edge-cloud and edge-only setups. Experiment results show that the proposed system can learn to detect a new object class in just a few minutes, including both dataset construction and model training. In comparison, traditional * Work was done when the author was a full-time employee with SRA. † Work was done when the author was a summer intern with SRA.fine-tuning based method may take a few hours for training, and in most cases would also need a tedious and costly manual dataset labeling step.

Continual Learning Approaches to Hand-Eye Calibration in Robots

“…In the class-incremental scenario, the model learns an increasing number of new classes using new data over time. An example of this scenario is learning new household objects for service robotics [31].…”

Republic of Korea (South Korea)

“…Two-phase setting. Table 1 shows that our method outperforms previous approaches such as LWF [29], RILOD [24], SID [39], and ERD [10] using GFLv1 [27], including CL-DETR [33]. Importantly, we achieved a 0.5% increase in AP for the 70+10 scenario and 1.0% for the 40+40 scenario.…”

“…Table 2 shows that our method, which utilizes synthetic image-based training, surprisingly outperforms other approaches significantly in multi-phase scenarios. Despite using different baselines like [10,24,39], it is evident that our method maintains consistent performance. We achieve 8.7% and 5.8% gains in AP for the 40+10+10+10+10 and 40+20+20 scenarios, respectively, compared to CL-DETR.…”

“…CIOD results (%) on COCO 2017 in two-phase setting. The results of related research [10,24,29,39] extract from CL-DETR paper. The order of data follows the [10].…”

“…Class incremental object detection. CIOD has progressed from primarily employing CNN-based methods [1,10,24,32,39,56] to also incorporating transformer-based approaches [14,33]. In this trend, ILOD [56], a pioneering work in CIOD, implemented the LWF [29] method to handle forgetting.…”