Counting in the Wild
In this paper we explore the scenario of learning to count multiple instances of objects from images that have been dot-annotated through crowdsourcing. Specifically, we work with a large and challenging image dataset of penguins in the wild, for which te
- PDF / 4,034,156 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 98 Downloads / 300 Views
2
Department of Engineering Science, University of Oxford, Oxford, UK [email protected] Skolkovo Institute of Science and Technology (Skoltech), Moscow, Russia
Abstract. In this paper we explore the scenario of learning to count multiple instances of objects from images that have been dot-annotated through crowdsourcing. Specifically, we work with a large and challenging image dataset of penguins in the wild, for which tens of thousands of volunteer annotators have placed dots on instances of penguins in tens of thousands of images. The dataset, introduced and released with this paper, shows such a high-degree of object occlusion and scale variation that individual object detection or simple counting-density estimation is not able to estimate the bird counts reliably. To address the challenging counting task, we augment and interleave density estimation with foreground-background segmentation and explicit local uncertainty estimation. The three tasks are solved jointly by a new deep multi-task architecture. Using this multi-task learning, we show that the spread between the annotators can provide hints about local object scale and aid the foreground-background segmentation, which can then be used to set a better target density for learning density prediction. Considerable improvements in counting accuracy over a single-task density estimation approach are observed in our experiments.
1
Introduction
This paper is motivated by the need to address a challenging large-scale realworld image-based counting problem that cannot be tackled well with existing approaches. This counting task arises in the course of ecological surveys of Antarctic penguins, and the images are automatically collected by a set of fixed cameras placed in Antarctica with the intention of monitoring the penguin population of the continent. The visual understanding of the collected images is compounded by many factors such as the variability of vantage points of the cameras, large variation of penguin scales, adversarial weather conditions in many images, high similarity of the appearance between the birds and some elements in the background (e.g. rocks), and extreme crowding and inter-occlusion between penguins (Fig. 1). The still ongoing annotation process of the dataset consists of a public website [27], where non-professional volunteers annotate images by placing dots on top of individual penguins; this is similar to citizen science annotators, who have also been used as an alternative to paid annotators for vision datasets (e.g. [19]). c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VII, LNCS 9911, pp. 483–498, 2016. DOI: 10.1007/978-3-319-46478-7 30
484
C. Arteta et al.
The simplest form of annotation (dotting) was chosen to scale up the annotation process as much as possible. Based on the large number of dot-annotated images, our goal is to train a deep model that can solve the counting task through density regression [4,6,8,16,25,26]. Compared to the training annotations used in previous w
Data Loading...
