Abstract: Training a Siamese architecture for re-identification with a large number of identities is a challenging task due to the difficulty of finding relevant negative samples efficiently. In this work we present Bag of Negatives (BoN), a method for accelerated and improved training of Siamese networks that scales well on datasets with a very large number of identities. BoN is an efficient and loss-independent method, able to select a bag of high quality negatives, based on a novel online hashing strategy.

Abstract: The increasing complexity learnt in the layers of a Convolutional Neural Network has proven to be of great help for the task of classification. The topic has received great attention in recently published literature.
Nonetheless, just a handful of works study low-level representations, commonly associated with lower layers. In this paper, we explore recent findings which conclude, counterintuitively, the last layer of the VGG convolutional network is the best to describe a low-level property such as texture. To shed some light on this issue, we are proposing a psychophysical experiment to evaluate the adequacy of different layers of the VGG network for texture retrieval. Results obtained suggest that, whereas the last convolutional layer is a good choice for a specific task of classification, it might not be the best choice as a texture descriptor, showing a very poor performance on texture retrieval. Intermediate layers show the best performance, showing a good combination of basic filters, as in the primary visual cortex, and also a degree of higher level information to describe more complex textures.

Abstract: Unsupervised representation learning aims at finding methods that learn representations from data without annotation-based signals. Abstaining from annotations not only leads to economic benefits but may – and to some extent already does – result in advantages regarding the representation’s structure, robustness, and generalizability to different tasks. In the long run, unsupervised methods are expected to surpass their supervised counterparts due to the reduction of human intervention and the inherently more general setup that does not bias the optimization towards an objective originating from specific annotation-based signals. While major advantages of unsupervised representation learning have been recently observed in natural language processing, supervised methods still dominate in vision domains for most tasks. In this dissertation, we contribute to the field of unsupervised (visual) representation learning from three perspectives: (i) Learning representations: We design unsupervised, backpropagation-free Convolutional Self-Organizing Neural Networks (CSNNs) that utilize self-organization- and Hebbian-based learning rules to learn convolutional kernels and masks to achieve deeper backpropagation-free models. Thereby, we observe that backpropagation-based and -free methods can suffer from an objective function mismatch between the unsupervised pretext task and the target task. This mismatch can lead to performance decreases for the target task. (ii) Evaluating representations: We build upon the widely used (non-)linear evaluation protocol to define pretext- and target-objective-independent metrics for measuring the objective function mismatch. With these metrics, we evaluate various pretext and target tasks and disclose dependencies of the objective function mismatch concerning different parts of the training and model setup. (iii) Transferring representations: We contribute CARLANE, the first 3-way sim-to-real domain adaptation benchmark for 2D lane detection. We adopt several well-known unsupervised domain adaptation methods as baselines and propose a method based on prototypical cross-domain self-supervised learning. Finally, we focus on pixel-based unsupervised domain adaptation and contribute a content-consistent unpaired image-to-image translation method that utilizes masks, global and local discriminators, and similarity sampling to mitigate content inconsistencies, as well as feature-attentive denormalization to fuse content-based statistics into the generator stream. In addition, we propose the cKVD metric to incorporate class-specific content inconsistencies into perceptual metrics for measuring translation quality.

Abstract: Few-shot learning has become essential for producing models that generalize from few examples. In this work, we identify that metric scaling and metric task conditioning are important to improve the performance of few-shot algorithms. Our analysis reveals that simple metric scaling completely changes the nature of few-shot algorithm parameter updates. Metric scaling provides improvements up to 14% in accuracy for certain metrics on the mini-Imagenet 5-way 5-shot classification task. We further propose a simple and effective way of conditioning a learner on the task sample set, resulting in learning a task-dependent metric space. Moreover, we propose and empirically test a practical end-to-end optimization procedure based on auxiliary task co-training to learn a task-dependent metric space. The resulting few-shot learning model based on the task-dependent scaled metric achieves state of the art on mini-Imagenet. We confirm these results on another few-shot dataset that we introduce in this paper based on CIFAR100.

Abstract: Colour constancy refers to the human visual system's ability to preserve the perceived colour of objects despite changes in the illumination. Its exact mechanisms are unknown, although a number of systems ranging from retinal to cortical and memory are thought to play important roles. The strength of the perceptual shift necessary to preserve these colours is usually estimated by the vectorial distances from an ideal match (or canonical illuminant). In this work we explore how much of the colour constancy phenomenon could be explained by well-known physiological properties of V1 and V2 neurons whose receptive fields (RF) vary according to the contrast and orientation of surround stimuli. Indeed, it has been shown that both RF size and the normalization occurring between centre and surround in cortical neurons depend on the local properties of surrounding stimuli. Our stating point is the construction of a computational model which includes this dynamical centre-surround adaptation by means of two overlapping asymmetric Gaussian kernels whose variances are adjusted to the contrast of surrounding pixels to represent the changes in RF size of cortical neurons and the weights of their respective contributions are altered according to differences in centre-surround contrast and orientation. The final output of the model is obtained after convolving an image with this dynamical operator and an estimation of the illuminant is obtained by considering the contrast of the far surround. We tested our algorithm on naturalistic stimuli from several benchmark datasets. Our results show that although our model does not require any training, its performance against the state-of-the-art is highly competitive, even outperforming learning-based algorithms in some cases. Indeed, these results are very encouraging if we consider that they were obtained with the same parameters for all datasets (i.e. just like the human visual system operates).

Abstract: We generated a new dataset of digital natural images where each colour plane corresponds to the human LMS (long-, medium-, short-wavelength) cone activations. The images were chosen to represent five different visual environments (eg forest, seaside, mountain snow, urban, motorways) and were taken under natural illumination at different times of day. At the bottom-left corner of each picture there was a matte grey ball of approximately constant spectral reflectance (across the camera's response spectrum,) and nearly Lambertian reflective properties, which allows to compute (and remove, if necessary) the illuminant's colour and intensity. The camera (Sigma Foveon SD10) was calibrated by measuring its sensor's spectral responses using a set of 31 spectrally narrowband interference filters. This allowed conversion of the final camera-dependent RGB colour space into the Smith and Pokorny (1975) cone activation space by means of a polynomial transformation, optimised for a set of 1269 Munsell chip reflectances. This new method is an improvement over the usual 3 × 3 matrix transformation which is only accurate for spectrally-narrowband colours. The camera-to-LMS transformation can be recalculated to consider other non-human visual systems. The dataset is available to download from our website.

Abstract: Color constancy (the ability to perceive colors relatively stable under different illuminants) is the result of several mechanisms spread across different neural levels and responding to several visual scene cues. It is usually measured by estimating the perceived color of a grey patch under an illuminant change. In this work, we hypothesize whether chromatic adaptation (without a reference white or grey) could be driven by certain colors, specifically those corresponding to the universal color terms proposed by Berlin and Kay (1969). To this end we have developed a new psychophysical paradigm in which subjects adjust the color of a test patch (in CIELab space) to match their memory of the best example of a given color chosen from the universal terms list (grey, red, green, blue, yellow, purple, pink, orange and brown). The test patch is embedded inside a Mondrian image and presented on a calibrated CRT screen inside a dark cabin. All subjects were trained to “recall” their most exemplary colors reliably from memory and asked to always produce the same basic colors when required under several adaptation conditions. These include achromatic and colored Mondrian backgrounds, under a simulated D65 illuminant and several colored illuminants. A set of basic colors were measured for each subject under neutral conditions (achromatic background and D65 illuminant) and used as “reference” for the rest of the experiment. The colors adjusted by the subjects in each adaptation condition were compared to the reference colors under the corresponding illuminant and a “constancy index” was obtained for each of them. Our results show that for some colors the constancy index was better than for grey. The set of best adapted colors in each condition were common to a majority of subjects and were dependent on the chromaticity of the illuminant and the chromatic background considered.