Page - 102 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

not successful. This happens when the tracking has “lost” the object and consecutive tracking iteration move the object around in the sensor data. In practice we have set this velocity to 80mm per frame. With 30 fps this means a velocity of about 2.4m per second. This is enough to accurately track quickly moving objects, while being robust to detect the random movements that typically occur when tracking of the object fails. 4. Experiments and Evaluation In order to compare the performance of our approach against the state of the art we use the synthetic dataset provided by Choi and Christensen [5]. The approach is also evaluated on real- world objects and the results are presented in [12]. Interested readers can find more information here2. The dataset in [5] consists of four object models and a synthetic test sequence (1000 RGB-D frames) for each object. The test sequence is obtained by placing each object in a virtual kitchen model and moving a virtual camera around the model. The object trajectories w.r.t the virtual camera coordinate frame serves as the ground truth pose (error free since it is generated via rendering) of the object. Fig. 2 shows one such frame of each object sequence. The performance of the proposed tracking approach on the synthetic data set is as shown in Figure 3. Figure 2. Example images from the synthetic test data set provided by [5], a) Milk b) Orange Juice c) Tide d) Kinect Box Figure 3: Results of the tracking approach on the synthetic data set a) Milk b) Orange Juice c) Tide and d) Kinect box 2 http://tracking.profactor.at 102