Week 8 Students: Meera Hahn and Si Chen Mentor: Afshin Deghan Cifar Network • 3 convolutional layer network • Trained Caffe classifier using this network and our training images from the first frames of the video sequence • Experimented with parameters such as number of training iterations, learning rate and number of training images Imagenet Network • 5 convolutional layers • This is the network we ran to train the offline Caffe tracker with Caffe’s given pre-trained weights which gave very high results • Trained Caffe classifier using this network and our training images from the first frames of the video sequence F Score Comparisons • Running Offline Caffe Deep Tracking code on all 50 sequences • Fixed: processing greyscale images • Calculated STRUCK results from benchmark results Autoencoder Fully Connected Offline CAFFE STRUCK + SVM Network Deep Tracker 115 Full 62.84 45.14 77.01 75.57 61.24 56.28 Online Object Tracking: A Benchmark 1 Spatial Robustness Evaluation (SRE) • 8 spatial shifts of the ground truth = 12 total shifts Precision Plot • Error in center location Temporal Robustness Evaluation (TRE) • Frames evaluated based on bounding box location Success Plot • Overlap in bounding boxes • Ratio of successful tracking Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang, “Online Object Tracking: A Benchmark,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 1