Large-Scale Training of Shadow Detectors with Noisily-Annotated Shadow Examples
This paper introduces training of shadow detectors under the large-scale dataset paradigm. This was previously impossible due to the high cost of precise shadow annotation. Instead, we advocate the use of quickly but imperfectly labeled images. Our novel
- PDF / 4,994,784 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 63 Downloads / 254 Views
Abstract. This paper introduces training of shadow detectors under the large-scale dataset paradigm. This was previously impossible due to the high cost of precise shadow annotation. Instead, we advocate the use of quickly but imperfectly labeled images. Our novel label recovery method automatically corrects a portion of the erroneous annotations such that the trained classifiers perform at state-of-the-art level. We apply our method to improve the accuracy of the labels of a new dataset that is 20 times larger than existing datasets and contains a large variety of scenes and image types. Naturally, such a large dataset is appropriate for training deep learning methods. Thus, we propose a semantic-aware patch level Convolutional Neural Network architecture that efficiently trains on patch level shadow examples while incorporating image level semantic information. This means that the detected shadow patches are refined based on image semantics. Our proposed pipeline can be a useful baseline for future advances in shadow detection.
Keywords: Shadow detection labels
1
·
Large scale shadow dataset
·
Noisy
Introduction
Shadows are ubiquitous in images of natural scenes. On one hand, shadows provide useful cues about the scene including object shapes [28], light sources and illumination conditions [23,30,31], camera parameters and geo-location [19], and scene geometry [21]. On the other hand, the presence of shadows in images creates difficulties for many computer vision tasks from image segmentation to object detection and tracking. In all cases, being able to automatically detect shadows, and subsequently remove them or reason about their shapes and sizes would usually be beneficial. Moreover, shadow-free images are of great interest for image editing, computational photography, and augmented reality, and the first crucial step is shadow detection. Shadow detection in single images is a well studied, but still challenging problem. Early work focused on physical modeling of the illumination and shadowing phenomena. Such approaches, e.g., illumination invariant methods [8,9], c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VI, LNCS 9910, pp. 816–832, 2016. DOI: 10.1007/978-3-319-46466-4 49
Large-Scale Training of Shadow Detectors with Noisily Annotated Examples
817
only work well for high quality images [24]. In contrast, for consumer-grade photographs and web quality images, the breakthrough in performance came with statistical learning approaches [12,17,24,49]. These approaches learn the appearance of shadows from images with ground-truth labels. The first sizable database with manually annotated shadows was the UCF shadow dataset [49], followed, soon after, by the UIUC shadow dataset [12]. These publicly available datasets with pixel-level annotations have led to important advances in the field. They enabled both systematic quantitative and qualitative evaluation of detection performance, as opposed to the prior practice of qualitative evaluation on a few selected images. In the
Data Loading...
