Seite - 101 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure 4: The problem of mul- tiple detections. Ground truth is shown in green. Left: state- of-the-art yields two bounding boxesofthesame,singleperson. Middle: two persons are visible. Detection yields two bounding boxes which are diffucult to as- sociate. Right: an even harder casewith three persons. Figure 5: This ROC plot shows results of Faster R- CNN(green),YOLO(blue),MaskR-CNN(red)and ourmethod (purple) for all occlusion levels. but also on the evaluation and on the detector which we leave open for future research. 5.Conclusion Thispaper formulatesanewscientificquestionon object detection with fragmented occlusion which is different to partial occlusion. We show by a study that current object detectors fail in this case. We generatedand labelledanewdataset showingpeople behind trees in a forestry environment. Such scenes frequentlyoccur inbordersurveillancewhichhasbe- come very important in EU security policies. We try to tackle the occlusion challenge by augmenting Mi- crosoftCOCOincludingthepixel-wisesegmentation masks to capture the occlusion problem. We show that Mask R-CNN trained on this data improves on fragmented occlusion, however, we also observe se- vere loss of spatial, structural information and that the bounding box itself is not the appropriate de- scription to cope with fragmented occlusion. This has severe implications on the detection approach it- self, but also on dataset labelling and evaluation. A potential solution is left open for futurework. Acknowledgments This research was supported by the European Union H2020 programme under grant agreement FOLDOUT-787021. We thank all our students on internship to label thenewdataset. References [1] M. Black and P. Anandan. The robust estimation of multiplemotions: Parametricandpiecewise-smooth flow fields. Computer Vision and Image Under- standing, 63:75–104,011996. [2] R.Girshick. Fast r-cnn. InCVPR,pages1440–1448, 2015. [3] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Richfeaturehierarchiesforaccurateobjectdetection and semantic segmentation. InCVPR, pages 580– 587,2014. [4] K.He,G.Gkioxari,P.Dolla´r,andR.Girshick. Mask r-cnn. InCVPR, pages2961–2969,2017. [5] S. Ioffe and C. Szegedy. Batch normaliza- tion: Accelerating deep network training by re- ducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. [6] W.Liu,D.Anguelov,D.Erhan,C.Szegedy,S.Reed, C.-Y.Fu, andA.C.Berg. Ssd: Single shotmultibox detector. InECCV, pages21–37.Springer, 2016. [7] G. Nebehay and R. Pflugfelder. Clustering of static-adaptive correspondences for deformable ob- ject tracking. InCVPR, June2015. [8] J. Redmon and A. Farhadi. Yolov3: An incremental improvement. arXiv, 2018. [9] S. Ren, K. He, R. Girshick, and J. Sun. Faster r- cnn: Towards real-time object detection with region proposalnetworks. InNIPS, pages 91–99,2015. [10] S.Ullman,L.Assif,E.Fetaya,andD.Harari. Atoms ofrecognitioninhumanandcomputervision.PNAS, 113(10):2744–2749,2016. 101