Crowdsourced 3D CAD models are becoming easily accessible online, and canpotentially generate an infinite number of training images for almost anyobject category.We show that augmenting the training data of contemporary DeepConvolutional Neural Net (DCNN) models with such synthetic data can beeffective, especially when real training data is limited or not well matched tothe target domain. Most freely available CAD models capture 3D shape but areoften missing other low level cues, such as realistic object texture, pose, orbackground. In a detailed analysis, we use synthetic CAD-rendered images toprobe the ability of DCNN to learn without these cues, with surprisingfindings. In particular, we show that when the DCNN is fine-tuned on the targetdetection task, it exhibits a large degree of invariance to missing low-levelcues, but, when pretrained on generic ImageNet classification, it learns betterwhen the low-level cues are simulated. We show that our synthetic DCNN trainingapproach significantly outperforms previous methods on the PASCAL VOC2007dataset when learning in the few-shot scenario and improves performance in adomain shift scenario on the Office benchmark.
展开▼
