Page - 96 - in Joint Austrian Computer Vision and Robotics Workshop 2020

tures is proposed in [3]. A recent work applies fine- tuning only region proposal and classification layers on a data set consisting of many base class and few newclass sampleswhilefixing the featureextraction partof thenetworkcanoutperformmeta-learningap- proaches [11]. 3.ProposedApproach We based our approach on [9], which we use as generic object detector and feature extractor. For the classification we follow the pipeline proposed in [12], which uses online random forests proposed in [8] as a classifier. The random forest can be incre- mentally trained, and is able to provide good results with few training samples. We use the model pre- trained on ImageNet from [9], and evaluate it on the 12 classes dataset provided with the authors’ imple- mentation1. Each of the classes has between 55 and 108 training samples. We compare to a linear clas- sifier trained on the entire set of samples, and train our online random forest based classifier with all or afixed subsetof samplesper class. With the full set of examples, the online random forest based classifier performs similarly but slightly worse than the linear classifier, with an F1 score of about 0.80. Down to about 20 samples per class, the performance stays nearly constant. With 10 samples the performance drops to around 0.70, with 5 sam- ples to about 0.67. Only then the performance starts to degrade more quickly, arriving at only about 1.5 times better than random when using a single sam- ple. The results are visualised in Figure 1. It is ap- parent, that the reduction of the F1-score is mainly due to reduced recall. In nearly all cases the loss in termsofrecall iscausedbymisclassifyingtheobject, while only in few cases the target object is missed in the detectionstage. 4.Conclusion Based on a recently proposed framework, which we use for generic object detection and feature ex- traction, we have developed an approach for few- shot object detection using an online random forest as a classifier, which makes it incrementally train- able. With about 20 samples there performance in terms of F1 score is similar to a linear classifier on thefull set, anddropsbyabout0.13whenusingonly 1https://github.com/mahyarnajibi/SNIPER/ tree/cvpr3k Figure 1. Detection results (F1 score, precision, recall) of theproposedapproachon the12classesdataset from[9], when trained on different numbers of samples per class. The confidence threshold is 0.15 for the online random forest classifier. 5 samples, which makes this a practically usable ap- proach inusecaseswith fewtrainingsamples. Acknowledgments This work has received funding from the Euro- pean Union’s Horizon 2020 research and innova- tion programme, under grant agreements n◦ 761802 MARCONI (“Multimedia and Augmented Radio Creation: Online, iNteractive, Individual”) and n◦ 761934, Hyper360 (“Enriching 360 media with 3D storytellingandpersonalisationelements”). References [1] H. Chen, Y. Wang, G. Wang, and Y. Qiao. Lstd: A low-shot transfer detector for object detection. In 32ndAAAIConf. onAI, 2018. [2] X. Dong, L. Zheng, F. Ma, Y. Yang, and D. Meng. Few-example object detection with model commu- nication. IEEE T.PAMI, 41(7), 2018. [3] B.Kang,Z.Liu,X.Wang,F.Yu,J.Feng,andT.Dar- rell. Few-shot object detection via feature reweight- ing. In Proc. ICCV, 2019. [4] L. Karlinsky, J. Shtok, S. Harary, E. Schwartz, A. Aides, R. Feris, R. Giryes, and A. M. Bronstein. Repmet: Representative-based metric learning for classificationandfew-shotobjectdetection. InProc. CVPR, 2019. [5] E.Maiettini,G.Pasquale,L.Rosasco,andL.Natale. Speeding-up object detection training for robotics with falkon. In RSJ IROS, 2018. [6] J. Redmon and A. Farhadi. Yolov3: An incremen- tal improvement. arXiv preprint arXiv:1804.02767, 2018. [7] S. Ren, K. He, R. Girshick, and J. Sun. Faster R- CNN: Towards real-time object detection with re- gion proposalnetworks. In NIPS, 2015. 96