Despite the success of hand-crafted features in computer visioning for many years, nowadays, this has been replaced by end-to-end learnable features that are extracted from deep convolutional neural networks (CNNs). Whilst CNNs can learn robust features directly from image pixels, they require large amounts of samples and extreme augmentations. On the contrary, hand-crafted features, like SIFT, exhibit several interesting properties as they can provide local rotation invariance. In this work, a novel scheme combining the strengths of SIFT descriptors with CNNs, namely SIFT-CNN, is presented. Given a single-channel image, one SIFT descriptor is computed for every pixel, and thus, every pixel is represented as an M-dimensional histogram, which ultimately results in an M-channel image. Thus, the SIFT image is generated from the SIFT descriptors for all the pixels in a single-channel image, while at the same time, the original spatial size is preserved. Next, a CNN is trained to utilize these M-channel images as inputs by operating directly on the multiscale SIFT images with the regular convolution processes. Since these images incorporate spatial relations between the histograms of the SIFT descriptors, the CNN is guided to learn features from local gradient information of images that otherwise can be neglected. In this manner, the SIFT-CNN implicitly acquires a local rotation invariance property, which is desired for problems where local areas within the image can be rotated without affecting the overall classification result of the respective image. Some of these problems refer to indirect immunofluorescence (IIF) cell image classification, ground-based all-sky image-cloud classification and human lip-reading classification. The results for the popular datasets related to the three different aforementioned problems indicate that the proposed SIFT-CNN can improve the performance and surpasses the corresponding CNNs trained directly on pixel values in various challenging tasks due to its robustness in local rotations. Our findings highlight the importance of the input image representation in the overall efficiency of a data-driven system.

Lip reading (LR) is the task of predicting the speech utilizing only the visual information of the speaker. In this work, for the first time, the benefits of alternating between spatiotemporal and spatial convolutions for learning effective features from the LR sequences are studied. In this context, a new learnable module named ALSOS (Alternating Spatiotemporal and Spatial Convolutions) is introduced in the proposed LR system. The ALSOS module consists of spatiotemporal (3D) and spatial (2D) convolutions along with two conversion components (3D-to-2D and 2D-to-3D) providing a sequence-to-sequence-mapping. The designed LR system utilizes the ALSOS module in-between ResNet blocks, as well as Temporal Convolutional Networks (TCNs) in the backend for classification. The whole framework is composed by feedforward convolutional along with residual layers and can be trained end-to-end directly from the image sequences in the word-level LR problem. The ALSOS module can capture spatiotemporal dynamics and can be advantageous in the task of LR when combined with the ResNet topology. Experiments with different combinations of ALSOS with ResNet are performed on a dataset in Greek language simulating a medical support application scenario and on the popular large-scale LRW-500 dataset of English words. Results indicate that the proposed ALSOS module can improve the performance of a LR system. Overall, the insertion of ALSOS module into the ResNet architecture obtained higher classification accuracy since it incorporates the contribution of the temporal information captured at different spatial scales of the framework.

In this work, we present a novel approach for training Generative Adversarial Networks (GANs). Using the attention maps produced by a Teacher- Network we are able to improve the quality of the generated images as well as perform weakly object localization on the generated images. To this end, we generate images of HEp-2 cells captured with Indirect Imunofluoresence (IIF) and study the ability of our network to perform a weakly localization of the cell. Firstly, we demonstrate that whilst GANs can learn the mapping between the input domain and the target distribution efficiently, the discriminator network is not able to detect the regions of interest. Secondly, we present a novel attention transfer mechanism which allows us to enforce the discriminator to put emphasis on the regions of interest via transfer learning. Thirdly, we show that this leads to more realistic images, as the discriminator learns to put emphasis on the area of interest. Fourthly, the proposed method allows one to generate both images as well as attention maps which can be useful for data annotation e.g in object detection.