The existing research on open-set recognition (OSR) and novel class discovery (NCD) has largely treated these tasks as independent fields. OSR aims to identify samples that do not belong to the training set classes, while NCD seeks to further classify such unseen, unlabeled samples into novel classes. However, there is a lack of a unified framework to automate both tasks systematically. In this paper, we propose a unified training framework to identify and categorize unseen samples. Specifically, we conduct a comprehensive evaluation of existing post hoc OSR methods and observe that their performance is highly sensitive to the temperature scaling factor. To address this, we introduce a distance-based evaluation method for OSR, which not only outperforms existing post hoc approaches but also integrates seamlessly with them to deliver enhanced performance. Furthermore, we developed a prototype-based classification head leveraging this distance metric, which facilitates compact feature representations for known classes and guides the clustering of unknown classes, thereby significantly enhancing the classification accuracy for novel classes. On the CUB-200-2011 dataset, our unified framework achieves a 0.95–6.12% improvement in AUROC scores on OSR benchmarks and a 3.19% increase in classification accuracy for novel classes. Extensive experiments and visualizations validate the effectiveness of the proposed approach. We believe that this unified framework will pave the way for automating the integration of OSR and NCD, offering a more efficient and systematic approach to addressing these tasks.

Lightweight robots are frequently engaged in real-time tracking tasks to provide human companionship services. For effective target tracking, the YOLO series is often employed as a lightweight object detection framework in robot systems. However, YOLO still demands substantial resources to train larger-scale models, striking a balance between accuracy and resource efficiency. Deploying YOLO directly on robots with limited computing resources can lead to significant delays in detection, compromising the effectiveness of tracking tasks. A deeper concern arises from the prevalent use of CPUs as the primary computing units in robots, rendering many existing model optimization techniques, which primarily target GPU computing, unsuitable for this context. To tackle this challenge, we propose a novel detection framework called SLCNet-YOLOv5, specifically designed for deployment in CPU-centric computing environments on robots. The core concept of SLCNet-YOLOv5 entails substituting the native YOLOv5 backbone network with SLCNet, which is a simplified version derived from the existing CPU convolutional neural network, PP-LCNet. It is important to note that our aim is not to enhance PP-LCNet to improve inference accuracy but rather to simplify it to enhance inference speed, while tolerating a certain degree of accuracy loss. This is because excessive inference latency may lead to real-time tracking failures. By employing a backbone network optimized for CPU-centric computation and reducing the computational complexity of the detection model, SLCNet markedly reduces latency, expediting the detection process, with only a minor trade-off in accuracy. In comparison to the performance of the state-of-the-art detector YOLOv5, experimental results on publicly available coco-foot-and-leg and PASCAL VOC datasets demonstrate significant enhancements in detection speed per image on CPU-centric terminals, with respective increases of 62.8% and 81.3%, alongside marginal declines in mean Average Precision (mAP) at 0.5 Intersection over Union (IoU) threshold, with losses of 0.077 and 0.165.