Abstract: Previous works on sequential learning address the problem of forgetting in discriminative models. In this paper we consider the case of generative models. In particular, we investigate generative adversarial networks (GANs) in the task of learning new categories in a sequential fashion. We first show that sequential fine tuning renders the network unable to properly generate images from previous categories (ie forgetting). Addressing this problem, we propose Memory Replay GANs (MeRGANs), a conditional GAN framework that integrates a memory replay generator. We study two methods to prevent forgetting by leveraging these replays, namely joint training with replay and replay alignment. Qualitative and quantitative experimental results in MNIST, SVHN and LSUN datasets show that our memory replay approach can generate competitive images while significantly mitigating the forgetting of previous categories.

Abstract: Siamese approaches address the visual tracking problem by extracting an appearance template from the current frame, which is used to localize the target in the next frame. In general, this template is linearly combined with the accumulated template from the previous frame, resulting in an exponential decay of information over time. While such an approach to updating has led to improved results, its simplicity limits the potential gain likely to be obtained by learning to update. Therefore, we propose to replace the handcrafted update function with a method which learns to update. We use a convolutional neural network, called UpdateNet, which given the initial template, the accumulated template and the template of the current frame aims to estimate the optimal template for the next frame. The UpdateNet is compact and can easily be integrated into existing Siamese trackers. We demonstrate the generality of the proposed approach by applying it to two Siamese trackers, SiamFC and DaSiamRPN. Extensive experiments on VOT2016, VOT2018, LaSOT, and TrackingNet datasets demonstrate that our UpdateNet effectively predicts the new target template, outperforming the standard linear update. On the large-scale TrackingNet dataset, our UpdateNet improves the results of DaSiamRPN with an absolute gain of 3.9% in terms of success score.

Abstract: Metric learning networks are used to compute image embeddings, which are widely used in many applications such as image retrieval and face recognition. In this paper, we propose to use network distillation to efficiently compute image embeddings with small networks. Network distillation has been successfully applied to improve image classification, but has hardly been explored for metric learning. To do so, we propose two new loss functions that model the
communication of a deep teacher network to a small student network. We evaluate our system in several datasets, including CUB-200-2011, Cars-196, Stanford Online Products and show that embeddings computed using small student networks perform significantly better than those computed using standard networks of similar size. Results on a very compact network (MobileNet-0.25), which can be
used on mobile devices, show that the proposed method can greatly improve Recall@1 results from 27.5% to 44.6%. Furthermore, we investigate various aspects of distillation for embeddings, including hint and attention layers, semisupervised learning and cross quality distillation.

Abstract: The cost of drawing object bounding boxes (ie labeling) for millions of images is prohibitively high. For instance, labeling pedestrians in a regular urban image could take 35 seconds on average. Active learning aims to reduce the cost of labeling by selecting only those images that are informative to improve the detection network accuracy. In this paper, we propose a method to perform active learning of object detectors based on convolutional neural networks. We propose a new image-level scoring process to rank unlabeled images for their automatic selection, which clearly outperforms classical scores. The proposed method can be applied to videos and sets of still images. In the former case, temporal selection rules can complement our scoring process. As a relevant use case, we extensively study the performance of our method on the task of pedestrian detection. Overall, the experiments show that the proposed method performs better than random selection.

Abstract: Mirror symmetry is a property most likely to be encountered in animals than in medium scale vegetation or inanimate objects in the natural world. This might be the reason why the human visual system has evolved to detect it quickly and robustly. Indeed, the perception of symmetry assists higher-level visual processing that are crucial for survival such as target recognition and identification irrespective of position and location. Although the task of detecting symmetrical objects seems effortless to us, it is very challenging for computers (to the extent that it has been proposed as a robust “captcha” by Funk & Liu in 2016). Indeed, the exact mechanism of symmetry detection in primates is not well understood: fMRI studies have shown that symmetrical shapes activate specific higher-level areas of the visual cortex (Sasaki et al.; 2005) and similarly, a large body of psychophysical experiments suggest that the symmetry perception is critically influenced by low-level mechanisms (Treder; 2010). In this work we attempt to find plausible low-level mechanisms that might form the basis for symmetry perception. Our simple model is made from banks of (i) odd-symmetric Gabors (resembling edge-detecting V1 neurons); and (ii) banks of larger odd- and even-symmetric Gabors (resembling higher visual cortex neurons), that pool signals from the 'edge image'. As reported previously (Akbarinia et al, ECVP2017), the convolution of the symmetrical lines with the two Gabor kernels of alternative phase produces a minimum in one and a maximum in the other (Osorio; 1996), and the rectification and combination of these signals create lines which hint of mirror symmetry in natural images. We improved the algorithm by combining these signals across several spatial scales. Our preliminary results suggest that such multiscale combination of convolutional operations might form the basis for much of the operation of the HVS in terms of symmetry detection and representation.

Abstract: In this study we present a unifed model of the visual cortex for predicting visual attention using real image scenes. Feedforward mechanisms from RGC and LGN have been functionally modeled using wavelet filters at distinct orientations and scales for each chromatic pathway (Magno-, Parvo-, Konio-cellular) and polarity (ON-/OFF-center), by processing image components in the CIE Lab space. In V1, we process cortical interactions with an excitatory-inhibitory network of fring rate neurons, initially proposed by (Li, 1999), later extended by (Penacchio et al. 2013). Firing rates from model’s output have been used as predictors of neuronal activity to be projected in a map in superior colliculus (with WTA-like computations), determining locations of visual fxations. These locations will be considered as already visited areas for future saccades, therefore we integrated a spatiotemporal function of inhibition of return mechanisms (where LIP/FEF is responsible) to feed to the model with spatial memory for next saccades. Foveation mechanisms have been simulated with a cortical magnifcation function, which distort spatial viewing properties for each fxation. Results show lower prediction errors than with respect no IoR cases (Fig. 1), and it is functionally consistent with human psychophysical measurements. Our model follows a biologically-constrained architecture, previously shown to reproduce visual saliency (Berga & Otazu, 2018), visual discomfort (Penacchio et al. 2016), brightness (Penacchio et al. 2013) and chromatic induction (Cerda & Otazu, 2016).

Abstract: A benchmark of saliency models performance with a synthetic image dataset is provided. Model performance is evaluated through saliency metrics as well as the influence of model inspiration and consistency with human psychophysics. SID4VAM is composed of 230 synthetic images, with known salient regions. Images were generated with 15 distinct types of low-level features (e.g. orientation, brightness, color, size...) with a target-distractor pop-out type of synthetic patterns. We have used Free-Viewing and Visual Search task instructions and 7 feature contrasts for each feature category. Our study reveals that state-of-the-art Deep Learning saliency models do not perform well with synthetic pattern images, instead, models with Spectral/Fourier inspiration outperform others in saliency metrics and are more consistent with human psychophysical experimentation. This study proposes a new way to evaluate saliency models in the forthcoming literature, accounting for synthetic images with uniquely low-level feature contexts, distinct from previous eye tracking image datasets.

Abstract: Recurrent neural networks (RNN) are popular for many computer vision tasks, including multi-label classification. Since RNNs produce sequential outputs, labels need to be ordered for the multi-label classification task. Current approaches sort labels according to their frequency, typically ordering them in either rare-first or frequent-first. These imposed orderings do not take into account that the natural order to generate the labels can change for each image, e.g.\ first the dominant object before summing up the smaller objects in the image. Therefore, in this paper, we propose ways to dynamically order the ground truth labels with the predicted label sequence. This allows for the faster training of more optimal LSTM models for multi-label classification. Analysis evidences that our method does not suffer from duplicate generation, something which is common for other models. Furthermore, it outperforms other CNN-RNN models, and we show that a standard architecture of an image encoder and language decoder trained with our proposed loss obtains the state-of-the-art results on the challenging MS-COCO, WIDER Attribute and PA-100K and competitive results on NUS-WIDE.

Abstract: One of the attractive characteristics of deep neural networks is their ability to transfer knowledge obtained in one domain to other related domains. As a result, high-quality networks can be trained in domains with relatively little training data. This property has been extensively studied for discriminative networks but has received significantly less attention for generative models. Given the often enormous effort required to train GANs, both computationally as well as in the dataset collection, the re-use of pretrained GANs is a desirable objective. We propose a novel knowledge transfer method for generative models based on mining the knowledge that is most beneficial to a specific target domain, either from a single or multiple pretrained GANs. This is done using a miner network that identifies which part of the generative distribution of each pretrained GAN outputs samples closest to the target domain. Mining effectively steers GAN sampling towards suitable regions of the latent space, which facilitates the posterior finetuning and avoids pathologies of other methods such as mode collapse and lack of flexibility. We perform experiments on several complex datasets using various GAN architectures (BigGAN, Progressive GAN) and show that the proposed method, called MineGAN, effectively transfers knowledge to domains with few target images, outperforming existing methods. In addition, MineGAN can successfully transfer knowledge from multiple pretrained GANs.