Description: Deep artificial neural networks have become the go-to method for many machine learning tasks. In the field of computer vision, deep convolutional neural networks achieve state-of-the-art performance for tasks such as classification, object detection, or instance segmentation. As deep neural networks become more and more complex, their inner workings become more and more opaque, rendering them a “black box” whose decision making process is no longer comprehensible. In recent years, various methods have been presented that attempt to peek inside the black box and to visualize the inner workings of deep neural networks, with a focus on deep convolutional neural networks for computer vision. These methods can serve as a toolbox to facilitate the design and inspection of neural networks for computer vision and the interpretation of the decision making process of the network. Here, we present the new tool Interactive Feature Localization in Deep neural networks (IFeaLiD) which provides a novel visualization approach to convolutional neural network layers. The tool interprets neural network layers as multivariate feature maps and visualizes the similarity between the feature vectors of individual pixels of an input image in a heat map display. The similarity display can reveal how the input image is perceived by different layers of the network and how the perception of one particular image region compares to the perception of the remaining image. IFeaLiD runs interactively in a web browser and can process even high resolution feature maps in real time by using GPU acceleration with WebGL 2. We present examples from four computer vision datasets with feature maps from different layers of a pre-trained ResNet101. IFeaLiD is open source and available online at https://ifealid.cebitec.uni-bielefeld.de.

Description: Marine researchers continue to create large quantities of benthic images e.g., using AUVs (Autonomous Underwater Vehicles). In order to quantify the size of sessile objects in the images, a pixel-to-centimeter ratio is required for each image, often indirectly provided through a geometric laser point (LP) pattern, projected onto the seafloor. Manual annotation of these LPs in all images is too time-consuming and thus infeasible for nowadays data volumes. Because of the technical evolution of camera rigs, the LP's geometrical layout and color features vary for different expeditions and projects. This makes the application of one algorithm, tuned to a strictly defined LP pattern, also ineffective. Here we present the web-tool DELPHI, that efficiently learns the LP layout for one image transect/collection from just a small number of hand labeled LPs and applies this layout model to the rest of the data. The efficiency in adapting to new data allows to compute the LPs and the pixel-to-centimeter ratio fully automatic and with high accuracy. DELPHI is applied to two real-world examples and shows clear improvements regarding reduction of tuning effort for new LP patterns as well as increasing detection performance.

Description: Marine imaging has evolved from small, narrowly focussed applications to large-scale applications covering areas of several hundred square kilometers or time series covering observation periods of several months. The analysis and interpretation of the accumulating large volume of digital images or videos will continue to challenge the marine science community to keep this process efficient and effective. It is safe to say that any strategy will rely on some software platform supporting manual image and video annotation, either for a direct manual annotation-based analysis or for collecting training data to deploy a machine learning–based approach for (semi-)automatic annotation. This paper describes how computer-assisted manual full-frame image and video annotation is currently performed in marine science and how it can evolve to keep up with the increasing demand for image and video annotation and the growing volume of imaging data. As an example, observations are presented how the image and video annotation tool BIIGLE 2.0 has been used by an international community of more than one thousand users in the last 4 years. In addition, new features and tools are presented to show how BIIGLE 2.0 has evolved over the same time period: video annotation, support for large images in the gigapixel range, machine learning assisted image annotation, improved mobility and affordability, application instance federation and enhanced label tree collaboration. The observations indicate that, despite novel concepts and tools introduced by BIIGLE 2.0, full-frame image and video annotation is still mostly done in the same way as two decades ago, where single users annotated subsets of image collections or single video frames with limited computational support. We encourage researchers to review their protocols for education and annotation, making use of newer technologies and tools to improve the efficiency and effectivity of image and video annotation in marine science.