Refine
Document type
- Conference Proceeding (11)
- Article (peer-reviewed) (2)
- Report (2)
Keywords
- 3D Imaging (1)
- CD-ROM (1)
- Classification (1)
- Convolutional neural networks (1)
- HTML5/WebGL (1)
- Image segmentation (1)
- Inertial sensing (1)
- Key frame selection (1)
- Real time (1)
- Semantic augmentation (1)
Cameras play a prominent role in the context of 3D data, as they can be designed to be very cheap and small and can therefore be used in many 3D reconstruction systems. Typical cameras capture video at 20 to 60 frames per second, resulting in a high number of frames to select from for 3D reconstruction. Many frames are unsuited for reconstruction as they suffer from motion blur or show too little variation compared to other frames. The camera used within this work has built-in inertial sensors. What if one could use the built-in inertial sensors to select a set of key frames well-suited for 3D reconstruction, free from motion blur and redundancy, in real time? A random forest classifier (RF) is trained by inertial data to determine frames without motion blur and to reduce redundancy. Frames are analyzed by the fast Fourier transformation and Lucas–Kanade method to detect motion blur and moving features in frames to label those correctly to train the RF. We achieve a classifier that omits successfully redundant frames and preserves frames with the required quality but exhibits an unsatisfied performance with respect to ideal frames. A 3D reconstruction by Meshroom shows a better result with selected key frames by the classifier. By extracting frames from video, one can comfortably scan objects and scenes without taking single pictures. Our proposed method automatically extracts the best frames in real time without using complex image-processing algorithms.
Collecting real-world data for the training of neural networks is enormously time-consuming and expensive. As such, the concept of virtualizing the domain and creating synthetic data has been analyzed in many instances. This virtualization offers many possibilities of changing the domain, and with that, enabling the relatively fast creation of data. It also offers the chance to enhance necessary augmentations with additional semantic information when compared with conventional augmentation methods. This raises the question of whether such semantic changes, which can be seen as augmentations of the virtual domain, contribute to better results for neural networks, when trained with data augmented this way. In this paper, a virtual dataset is presented, including semantic augmentations and automatically generated annotations, as well as a comparison between semantic and conventional augmentation for image data. It is determined that the results differ only marginally for neural network models trained with the two augmentation approaches.