Bezier Line Object Detection
Rostock, Univ., Master Thesis, 2021
Detecting small and thin objects is the most challenging part of object detection in general. The novel object detectors today contribute heavily in detecting generic objects that are mostly big and visible making it easier for the detectors to extract features and creating bounding boxes over them. Those detectors often fail in detecting objects in specialized applications that are either very small or thin such as lines/curves. Existing approaches to detect lines are mostly concentrated in detecting polylines that are often hard to annotate, uses more computation, formulated in recurrent ways. To solve this, our algorithm suggests a generic solution to detect bezier lines using a single-shot object detection approach. Instead of detecting individual polylines, our algorithm suggests a formulation to detect individual lines using their very control points. This method has several advantages over previous methods. It does not just create lines instead of boxes but detect the coordinates of the individual lines as a whole with good accuracy. Comparable to polylines, the targets are easily annotatable and use less computation power since require only control point coordinates to generate accurate results. We evaluated our results on the TUSimple lane dataset and compared the results with other line detectors with satisfactory results, hereby demonstrating the ability of this algorithm to work on generalized datasets as well. The source code is available on Github.
Identification and Measurement of Individual Roots in Minirhizotron Images of Dense Root Systems
2021 International Conference on Computer Vision Workshops. Proceedings
International Conference on Computer Vision Workshops (ICCVW) <2021, Online>
Semantic segmentation networks are prone to oversegmentation in areas where objects are tightly clustered. In minirhizotron images with densely packed plant root systems this can lead to a failure to separate individual roots, thereby skewing the root length and width measurements. We propose to deal with this problem by adding additional output heads to the segmentation model, one of which is used with a ridge detection algorithm as an intermediate step and a second one that directly estimates root width. With this method we are able to improve detection and width measurements in densely packed roots systems without negative effects on sparse root systems.
Mask, Train, Repeat! Artificial Intelligence for Quantitative Wood Anatomy
Frontiers in Plant Science
The recent developments in artificial intelligence have the potential to facilitate new research methods in ecology. Especially Deep Convolutional Neural Networks (DCNNs) have been shown to outperform other approaches in automatic image analyses. Here we apply a DCNN to facilitate quantitative wood anatomical (QWA) analyses, where the main challenges reside in the detection of a high number of cells, in the intrinsic variability of wood anatomical features, and in the sample quality. To properly classify and interpret features within the images, DCNNs need to undergo a training stage. We performed the training with images from transversal wood anatomical sections, together with manually created optimal outputs of the target cell areas. The target species included an example for the most common wood anatomical structures: four conifer species; a diffuse-porous species, black alder (Alnus glutinosa L.); a diffuse to semi-diffuse-porous species, European beech (Fagus sylvatica L.); and a ring-porous species, sessile oak (Quercus petraea Liebl.). The DCNN was created in Python with Pytorch, and relies on a Mask-RCNN architecture. The developed algorithm detects and segments cells, and provides information on the measurement accuracy. To evaluate the performance of this tool we compared our Mask-RCNN outputs with U-Net, a model architecture employed in a similar study, and with ROXAS, a program based on traditional image analysis techniques. First, we evaluated how many target cells were correctly recognized. Next, we assessed the cell measurement accuracy by evaluating the number of pixels that were correctly assigned to each target cell. Overall, the “learning process” defining artificial intelligence plays a key role in overcoming the issues that are usually manually solved in QWA analyses. Mask-RCNN is the model that better detects which are the features characterizing a target cell when these issues occur. In general, U-Net did not attain the other algorithms’ performance, while ROXAS performed best for conifers, and Mask-RCNN showed the highest accuracy in detecting target cells and segmenting lumen areas of angiosperms. Our research demonstrates that future software tools for QWA analyses would greatly benefit from using DCNNs, saving time during the analysis phase, and providing a flexible approach that allows model retraining.
Out of Distribution Detection for Generalized Zero Shot Learning
Rostock, Univ., Master Thesis, 2021
Zero Shot Learning (ZSL) is a problem set which classifies instances of the class samples not seen during training. Taking into account a more realistic, and real world scenario of classification, a Generalized Zero Shot Learning (GZSL) setting, in which the testing is made more complex by adding training class samples to a pool of unseen class samples. GZSL models are mostly designed using deep learning framework, but deep learning models designed using neural networks are biased towards the training class data and try to classify maximum instances as a seen classes. When deployed in real world applications, the Out of Distribution (OOD) data can be seen frequently by these models. Neural networks are inexperienced in managing an OOD data, can give incorrect prediction with a high confidence about them, or classify them into the positive classes. Therefore Outof- Distribution detection in neural networks is a crucial aspect in a model training and its deployment in real world task, to keep model’s sanity. In this thesis we designed an attribute detector model to predict attributes of an input image and tried to classify seen and unseen data efficiently into respective classes with foreseeing OOD instances. We evaluated several methods for detection of an OOD input for our designed model, and classification classification of the GZSL test data. After carrying out the experiments, we found out that, In Distribution (InD) and OOD images can be detected and separated successfully, without hampering the accuracy of our designed model.
Towards Combined Open Set Recognition and Out-of-Distribution Detection for Fine-grained Classification
Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications
International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP) <16, 2021, online>
We analyze the two very similar problems of Out-of-Distribution (OOD) Detection and Open Set Recognition (OSR) in the context of fine-grained classification. Both problems are about detecting object classes that a classifier was not trained on, but while the former aims to reject invalid inputs, the latter aims to detect valid but unknown classes. Previous works on OOD detection and OSR methods are evaluated mostly on very simple datasets or datasets with large inter-class variance and perform poorly in the fine-grained setting. In our experiments, we show that object detection works well to recognize invalid inputs and techniques from the field of fine-grained classification, like individual part detection or zooming into discriminative local regions, are helpful for fine-grained OSR.