On Pre-Trained Image Features and Synthetic Images for Deep Learning

Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, Kurt Konolige; Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0-0

DeepLearningmethodsusuallyrequirehugeamountsoftrainingdata to perform at their full potential, and often require expensive manual labeling. Using synthetic images is therefore very attractive to train object detectors, as the labeling comes for free, and several approaches have been proposed to combine synthetic and real images for training. In this paper, we evaluate if ’freezing’ the layers responsible for feature extraction to generic layers pre-trained on real images, and training only the remaining layers with plain OpenGL rendering may allow for training with synthetic images only. Our experiments with very recent deep architectures for object recognition (Faster-RCNN, R-FCN, Mask-RCNN) and image feature extractors (InceptionResnet and Resnet) show this simple approach performs surprisingly well.