A Large Contextual Dataset for Classification, Detection and Counting of Cars with Deep Learning
We have created a large diverse set of cars from overhead images (Data sets, annotations, networks and scripts are available from http://gdo-datasci.ucllnl.org/cowc/ ), which are useful for training a deep learner to binary classify, detect and count them
- PDF / 5,050,264 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 48 Downloads / 242 Views
Abstract. We have created a large diverse set of cars from overhead images (Data sets, annotations, networks and scripts are available from http://gdo-datasci.ucllnl.org/cowc/), which are useful for training a deep learner to binary classify, detect and count them. The dataset and all related material will be made publically available. The set contains contextual matter to aid in identification of difficult targets. We demonstrate classification and detection on this dataset using a neural network we call ResCeption. This network combines residual learning with Inceptionstyle layers and is used to count cars in one look. This is a new way to count objects rather than by localization or density estimation. It is fairly accurate, fast and easy to implement. Additionally, the counting method is not car or scene specific. It would be easy to train this method to count other kinds of objects and counting over new scenes requires no extra set up or assumptions about object locations. Keywords: Deep · Learning · CNN Automobile · Classification · Detection
1
· COWC · · Counting
Context
·
Cars
·
Introduction
Automated analytics involving detection, tracking and counting of automobiles from satellite or aerial platform are useful for both commercial and government purposes. For instance, [1] have developed a product to count cars in parking lots for investment customers who wish to monitor the business volume of retailers. Governments can also use tracking and counting data to monitor volume and pattern of traffic as well as volume of parking. If satellite data is cheap and plentiful enough, then it can be more cost effective than embedding sensors in the road. A problem encountered when trying to create automated systems for these purposes is a lack of large standardized public datasets. For instance OIRDS [2] has only 180 unique cars. A newer set VEDAI [3] has 2950 cars. However, both of these datasets are limited by not only the number of unique objects, but they c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part III, LNCS 9907, pp. 785–800, 2016. DOI: 10.1007/978-3-319-46487-9 48
786
T.N. Mundhenk et al.
also tend to cover the same region or use the same sensors. For instance, all images in the VEDAI set come from the AGRC Utah image collection [4]. We have created a new large dataset of Cars Overhead with Context (COWC). Our set contains a large number of unique cars (32,716) from six different image sets each covering a different geographical location and produced by different imagers. The images cover regions from Toronto Canada [5], Selwyn New Zealand [6], Potsdam [7] and Vaihingen Germany [8], Columbus [9] and Utah [4] United States. The set is also designed to be difficult. It contains 58,247 usable negative targets. Many of these have been hand picked from items easy to mistake for cars. Examples of these are boats, trailers, bushes and A/C units. To compensate for the added difficulty, context is included around targets. Context can help tell us something may not be a car (is sitting
Data Loading...
