Single-target tracking of generic objects is a difficult task since a trained tracker is given information present only in the first frame of a video. In recent years, increasingly many trackers have been based on deep neural networks that learn generic features relevant for tracking. This paper argues that deep architectures are often fit to learn implicit representations of optical flow. Optical flow is intuitively useful for tracking, but most deep trackers must learn it implicitly. This paper is among the first to study the role of optical flow in deep visual tracking. The architecture of a typical tracker is modified to reveal the presence of implicit representations of optical flow and to assess the effect of using the flow information more explicitly. The results show that the considered network learns implicitly an effective representation of optical flow. The implicit representation can be replaced by an explicit flow input without a notable effect on performance. Using the implicit and explicit representations at the same time does not improve tracking accuracy. The explicit flow input could allow constructing lighter networks for tracking.

Abstract-An approach to automatically detect and classify young spruce and birch trees in forest environment is presented. The method could be used in autonomous or semi-autonomous forest machines during tending operations. Detection is done by segmenting laser range images formed by a rotating laser scanner. Classification is done with a two-class Naive Bayes classifier based on image texture features. Multiple combinations of 99 features were tested and the best classifier included eight features from the co-occurrence matrix, local binary patterns, statistical geometrical features and Gabor filter. 79% of spruces and birches in the testing material were detected and 74% of these were correctly classified. Results suggest that the approach is suitable but there are still some challenges in each of the processing steps. Iteration between segmentation and classification is needed to increase reliability.

Log identification is an important task in silviculture and forestry. It involves matching tree logs with each other and telling which of the known individuals a given specimen is. Forest harvesters can image the logs and assess their quality while cutting trees in the forest. Identification allows each log to be traced back to the location it was grown in and efficiently choosing logs of specific quality in the sawmill. In this paper, a deep two-stream convolutional neural network is used to measure the likelihood that a pair of images represents the same part of a log. The similarity between the images is assessed based on the cross-correlation of the convolutional feature maps at one or more levels of the network. The performance of the network is evaluated with two large datasets, containing either spruce or pine logs. The best architecture identifies correctly 99% of the test logs in the spruce dataset and 97% of the test logs in the pine dataset. The results show that the proposed model performs very well in relatively good conditions. The analysis forms a basis for future attempts to utilize deep networks for log identification in challenging real-world forestry applications.