We address the problem of speech denoising where the goal is to extract clean speech signal from a noisy signal. Traditionally, the task of denoising has been performed using audio modality only. However, human speech perception is inherently multimodal where cues from visual modality are used to understand the speech better in a noisy environment. Similar observation has been made with computational denoising methods, i.e., performance of audio only model improves after adding visual modality. Inspired by these findings we propose a novel audio-visual network for adaptively combining both modalities for the task of speech audio denoising. We show that extracting noise from mixed audio and using it as a conditioning signal, improves speech denoising performance. To estimate the noise, we use both audio and visual modalities, i.e., lip region of the speaker, to extract the non-speech/silent regions from it. The silent regions enable us to estimate better noise profile to eliminate from the signal. Our proposed network uses self and cross attention framework between audio and video features, along the temporal dimension, to model correlations between the two modalities. We evaluate the proposed approach on a large scale audio-visual dataset VoxCeleb2 and obtain state-of-the-art results. We also demonstrate generalization to unseen speakers at test time.

We introduce \textit{Preserve Anything}, a novel method for controlled image synthesis that addresses key limitations in object preservation and semantic consistency in text-to-image (T2I) generation. Existing approaches often fail (i) to preserve multiple objects with fidelity, (ii) maintain semantic alignment with prompts, or (iii) provide explicit control over scene composition. To overcome these challenges, the proposed method employs an N-channel ControlNet that integrates (i) object preservation with size and placement agnosticism, color and detail retention, and artifact elimination, (ii) high-resolution, semantically consistent backgrounds with accurate shadows, lighting, and prompt adherence, and (iii) explicit user control over background layouts and lighting conditions. Key components of our framework include object preservation and background guidance modules, enforcing lighting consistency and a high-frequency overlay module to retain fine details while mitigating unwanted artifacts. We introduce a benchmark dataset consisting of 240K natural images filtered for aesthetic quality and 18K 3D-rendered synthetic images with metadata such as lighting, camera angles, and object relationships. This dataset addresses the deficiencies of existing benchmarks and allows a complete evaluation. Empirical results demonstrate that our method achieves state-of-the-art performance, significantly improving feature-space fidelity (FID 15.26) and semantic alignment (CLIP-S 32.85) while maintaining competitive aesthetic quality. We also conducted a user study to demonstrate the efficacy of the proposed work on unseen benchmark and observed a remarkable improvement of $\sim25\%$, $\sim19\%$, $\sim13\%$, and $\sim14\%$ in terms of prompt alignment, photorealism, the presence of AI artifacts, and natural aesthetics over existing works.

We present a novel multimodal multitask network and associated training algorithm. The method is capable of ingesting data from approximately 12 different modalities namely image, video, audio, text, depth, point cloud, time series, tabular, graph, X-ray, infrared, IMU, and hyper-spectral. The proposed approach utilizes modality specialized tokenizers, a shared transformer architecture, and cross-attention mechanisms to project the data from different modalities into a unified embedding space. It addresses multimodal and multitask scenarios by incorpo-rating modality-specific task heads for different tasks in respective modalities. We propose a novel pretraining strategy with iterative modality switching to initialize the network, and a training algorithm which trades off fully joint training over all modalities, with training on pairs of modalities at a time. We provide comprehensive evaluation across 25 datasets from 12 modalities and show state of the art performances, demonstrating the effectiveness of the proposed architecture, pretraining strategy and adapted multitask training.

Majority of research in learning based methods has been towards designing and training networks for specific tasks. However, many of the learning based tasks, across modalities, share commonalities and could be potentially tackled in a joint framework. We present an approach in such direction, to learn multiple tasks, in multiple modalities, with a unified architecture. The proposed network is composed of task specific encoders, a common trunk in the middle, followed by task specific prediction heads. We first pre-train it by self-supervised masked training, followed by sequential training for the different tasks. We train the network on all major modalities, e.g. visual, audio, text and 3D, and report results on 22 diverse and challenging public benchmarks. We demonstrate empirically that, using a joint network to train across modalities leads to meaningful information sharing and this allows us to achieve state-of-the-art results on most of the benchmarks. We also show generalization of the trained network on cross-modal tasks as well as unseen datasets and tasks.

Binaural audio gives the listener an immersive experience and can enhance augmented and virtual reality. However, recording binaural audio requires specialized setup with a dummy human head having microphones in left and right ears. Such a recording setup is difficult to build and setup, therefore mono audio has become the preferred choice in common devices. To obtain the same impact as binaural audio, recent efforts have been directed towards lifting mono audio to binaural audio conditioned on the visual input from the scene. Such approaches have not used an important cue for the task: the distance of different sound producing objects from the microphones. In this work, we argue that depth map of the scene can act as a proxy for inducing distance information of different objects in the scene, for the task of audio binauralization. We propose a novel encoder-decoder architecture with a hierarchical attention mechanism to encode image, depth and audio feature jointly. We design the network on top of state-of-the-art transformer networks for image and depth representation. We show empirically that the proposed method outperforms state-of-the-art methods comfortably for two challenging public datasets FAIR-Play and MUSIC-Sereo. We also demonstrate with qualitative results that the method is able to focus on the right information required for the task. The qualitative results are available at our project page https://krantiparida.github.io/projects/bmonobinaural.html

We address the problem of estimating depth with multi modal audio visual data. Inspired by the ability of animals, such as bats and dolphins, to infer distance of objects with echolocation, some recent methods have utilized echoes for depth estimation. We propose an end-to-end deep learning based pipeline utilizing RGB images, binaural echoes and estimated material properties of various objects within a scene. We argue that the relation between image, echoes and depth, for different scene elements, is greatly influenced by the properties of those elements, and a method designed to leverage this information can lead to significantly improved depth estimation from audio visual inputs. We propose a novel multi modal fusion technique, which incorporates the material properties explicitly while combining audio (echoes) and visual modalities to predict the scene depth. We show empirically, with experiments on Replica dataset, that the proposed method obtains 28% improvement in RMSE compared to the state-of-the-art audio-visual depth prediction method. To demonstrate the effectiveness of our method on larger dataset, we report competitive performance on Matterport3D, proposing to use it as a multimodal depth prediction benchmark with echoes for the first time. We also analyse the proposed method with exhaustive ablation experiments and qualitative results. The code and models are available at https://krantiparida.github.io/projects/bimgdepth.html

We propose simple yet effective improvements in point representations and local neighborhood graph construction within the general framework of graph neural networks (GNNs) for 3D point cloud processing. As a first contribution, we propose to augment the vertex representations with important local geometric information of the points, followed by nonlinear projection using a MLP. As a second contribution, we propose to improve the graph construction for GNNs for 3D point clouds. The existing methods work with a k-NN based approach for constructing the local neighborhood graph. We argue that it might lead to reduction in coverage in case of dense sampling by sensors in some regions of the scene. The proposed methods aims to counter such problems and improve coverage in such cases. As the traditional GNNs were designed to work with general graphs, where vertices may have no geometric interpretations, we see both our proposals as augmenting the general graphs to incorporate the geometric nature of 3D point clouds. While being simple, we demonstrate with multiple challenging benchmarks, with relatively clean CAD models, as well as with real world noisy scans, that the proposed method achieves state of the art results on benchmarks for 3D classification (ModelNet40) , part segmentation (ShapeNet) and semantic segmentation (Stanford 3D Indoor Scenes Dataset). We also show that the proposed network achieves faster training convergence, i.e. ∼ 40% less epochs for classification. The project details are available at https://siddharthsrivastava.github.io/publication/geomgcnn/

We address the problem of completing per pixel dense depth map using a single RGB image and the sparse point cloud of the scene. Depth prediction from RGB image is a hard problem and while dense point clouds obtained from LiDAR sensors can be used in addition to RGB image, the cost of such sensors is a significant barrier. Having LiDAR sensors which capture sparse point clouds is a reasonable middle ground. We propose a novel architecture which incorporates geometric primitives and self attention mechanisms, to improve the prediction. The motivation of self attention is to capture the correlations between scene and object elements, e.g. between the right and left window of car, early on in the network. While that for using geometric primitives is to have a high level clustering cue to enable the network to exploit similar correlations. In addition, we enforce complimentarity in the predictions made with RGB and sparse LiDAR respectively, this forces the two corresponding branches to focus on hard areas which are not already well predicted by the other branch. With exhaustive experiments on KITTI depth completion benchmark, NYU v2 and Matterport3D we show that the proposed method provides state-of-the-art results.

We address the problem of 3D object detection from 2D monocular images in autonomous driving scenarios. We propose to lift the 2D images to 3D representations using learned neural networks and leverage existing networks working directly on 3D data to perform 3D object detection and localization. We show that, with carefully designed training mechanism and automatically selected minimally noisy data, such a method is not only feasible, but gives higher results than many methods working on actual 3D inputs acquired from physical sensors. On the challenging KITTI benchmark, we show that our 2D to 3D lifted method outperforms many recent competitive 3D networks while significantly outperforming previous state-of-the-art for 3D detection from monocular images. We also show that a late fusion of the output of the network trained on generated 3D images, with that trained on real 3D images, improves performance. We find the results very interesting and argue that such a method could serve as a highly reliable backup in case of malfunction of expensive 3D sensors, if not potentially making them redundant, at least in the case of low human injury risk autonomous navigation scenarios like warehouse automation.

We present an approach for unsupervised training of CNNs in order to learn discriminative face representations. We mine supervised training data by noting that multiple faces in the same video frame must belong to different persons and the same face tracked across multiple frames must belong to the same person. We obtain millions of face pairs from hundreds of videos without using any manual supervision. Although faces extracted from videos have a lower spatial resolution than those which are available as part of standard supervised face datasets such as LFW and CASIA-WebFace, the former represent a much more realistic setting, e.g. in surveillance scenarios where most of the faces detected are very small. We train our CNNs with the relatively low resolution faces extracted from video frames collected, and achieve a higher verification accuracy on the benchmark LFW dataset cf. hand-crafted features such as LBPs, and even surpasses the performance of state-of-the-art deep networks such as VGG-Face, when they are made to work with low resolution input images.

We present a method for discovering never-seen-before objects in 3D point clouds obtained from sensors like Microsoft Kinect. We generate supervoxels directly from the point cloud data and use them with a Siamese network, built on a recently proposed 3D convolutional neural network architecture. We use known objects to train a non-linear embedding of supervoxels, by optimizing the criteria that supervoxels which fall on the same object should be closer than those which fall on different objects, in the embedding space. We test on unknown objects, which were not seen during training, and perform clustering in the learned embedding space of supervoxels to effectively perform novel object discovery. We validate the method with extensive experiments, quantitatively showing that it can discover numerous unseen objects while being trained on only a few dense 3D models. We also show very good qualitative results of object discovery in point cloud data when the test objects, either specific instances or even categories, were never seen during training.

We address the problem of video classification for facial analysis and human action recognition. We propose a novel weakly supervised learning method that models the video as a sequence of automatically mined, discriminative sub-events (e.g., onset and offset phase for "smile", running and jumping for "highjump"). The proposed model is inspired by the recent works on Multiple Instance Learning and latent SVM/HCRF - it extends such frameworks to model the ordinal aspect in the videos, approximately. We obtain consistent improvements over relevant competitive baselines on four challenging and publicly available video based facial analysis datasets for prediction of expression, clinical pain and intent in dyadic conversations, and on three challenging human action datasets. We also validate the method with qualitative results and show that they largely support the intuitions behind the method.

We propose a novel method for temporally pooling frames in a video for the task of human action recognition. The method is motivated by the observation that there are only a small number of frames which, together, contain sufficient information to discriminate an action class present in a video, from the rest. The proposed method learns to pool such discriminative and informative frames, while discarding a majority of the non-informative frames in a single temporal scan of the video. Our algorithm does so by continuously predicting the discriminative importance of each video frame and subsequently pooling them in a deep learning framework. We show the effectiveness of our proposed pooling method on standard benchmarks where it consistently improves on baseline pooling methods, with both RGB and optical flow based Convolutional networks. Further, in combination with complementary video representations, we show results that are competitive with respect to the state-of-the-art results on two challenging and publicly available benchmark datasets.

We introduce an Expanded Parts Model (EPM) for recognizing human attributes (e.g., young, short hair, wearing suits) and actions (e.g., running, jumping) in still images. An EPM is a collection of part templates which are learnt discriminatively to explain specific scale-space regions in the images (in human centric coordinates). This is in contrast to current models which consist of a relatively few (i.e., a mixture of) 'average' templates. EPM uses only a subset of the parts to score an image and scores the image sparsely in space, i.e., it ignores redundant and random background in an image. To learn our model, we propose an algorithm which automatically mines parts and learns corresponding discriminative templates together with their respective locations from a large number of candidate parts. We validate our method on three recent challenging datasets of human attributes and actions. We obtain convincing qualitative and state-of-the-art quantitative results on the three datasets.

We propose a novel Coupled Projection multi-task Metric Learning (CP-mtML) method for large scale face retrieval. In contrast to previous works which were limited to low dimensional features and small datasets, the proposed method scales to large datasets with high dimensional face descriptors. It utilises pairwise (dis-)similarity constraints as supervision and hence does not require exhaustive class annotation for every training image. While, traditionally, multi-task learning methods have been validated on same dataset but different tasks, we work on the more challenging setting with heterogeneous datasets and different tasks. We show empirical validation on multiple face image datasets of different facial traits, e.g. identity, age and expression. We use classic Local Binary Pattern (LBP) descriptors along with the recent Deep Convolutional Neural Network (CNN) features. The experiments clearly demonstrate the scalability and improved performance of the proposed method on the tasks of identity and age based face image retrieval compared to competitive existing methods, on the standard datasets and with the presence of a million distractor face images.

We study the problem of facial analysis in videos. We propose a novel weakly supervised learning method that models the video event (expression, pain etc.) as a sequence of automatically mined, discriminative sub-events (e.g. onset and offset phase for smile, brow lower and cheek raise for pain). The proposed model is inspired by the recent works on Multiple Instance Learning and latent SVM/HCRF - it extends such frameworks to model the ordinal or temporal aspect in the videos, approximately. We obtain consistent improvements over relevant competitive baselines on four challenging and publicly available video based facial analysis datasets for prediction of expression, clinical pain and intent in dyadic conversations. In combination with complimentary features, we report state-of-the-art results on these datasets.

We present a novel latent embedding model for learning a compatibility function between image and class embeddings, in the context of zero-shot classification. The proposed method augments the state-of-the-art bilinear compatibility model by incorporating latent variables. Instead of learning a single bilinear map, it learns a collection of maps with the selection, of which map to use, being a latent variable for the current image-class pair. We train the model with a ranking based objective function which penalizes incorrect rankings of the true class for a given image. We empirically demonstrate that our model improves the state-of-the-art for various class embeddings consistently on three challenging publicly available datasets for the zero-shot setting. Moreover, our method leads to visually highly interpretable results with clear clusters of different fine-grained object properties that correspond to different latent variable maps.

We propose a novel joint learning method for cross domain age estimation, a domain adaptation problem. The proposed method learns a low dimensional projection along with a re-gressor, in the projection space, in a joint framework. The projection aligns the features from two different domains, i.e. source and target, to the same space, while the regressor predicts the age from the domain aligned features. After this alignment, a regressor trained with only a few examples from the target domain, along with more examples from the source domain, can predict very well the ages of the target domain face images. We provide empirical validation on the largest publicly available dataset for age estimation i.e. MORPH-II. The proposed method improves performance over several strong baselines and the current state-of-the-art methods.

We propose a novel algorithm for the task of supervised discriminative distance learning by nonlinearly embedding vectors into a low dimensional Euclidean space. We work in the challenging setting where supervision is with constraints on similar and dissimilar pairs while training. The proposed method is derived by an approximate kernelization of a linear Mahalanobis-like distance metric learning algorithm and can also be seen as a kernel neural network. The number of model parameters and test time evaluation complexity of the proposed method are O(dD) where D is the dimensionality of the input features and d is the dimension of the projection space -- this is in contrast to the usual kernelization methods as, unlike them, the complexity does not scale linearly with the number of training examples. We propose a stochastic gradient based learning algorithm which makes the method scalable (w.r.t. the number of training examples), while being nonlinear. We train the method with up to half a million training pairs of 4096 dimensional CNN features. We give empirical comparisons with relevant baselines on seven challenging datasets for the task of low dimensional semantic category based image retrieval.

Retrieving images for an arbitrary user query, provided in textual form, is a challenging problem. A recently proposed method addresses this by constructing a visual classifier with images returned by an internet image search engine, based on the user query, as positive images while using a fixed pool of negative images. However, in practice, not all the images obtained from internet image search are always pertinent to the query; some might contain abstract or artistic representation of the content and some might have artifacts. Such images degrade the performance of on-the-fly constructed classifier. We propose a method for improving the performance of on-the-fly classifiers by using transfer learning via attributes. We first map the textual query to a set of known attributes and then use those attributes to prune the set of images downloaded from the internet. This pruning step can be seen as zero-shot learning of the visual classifier for the textual user query, which transfers knowledge from the attribute domain to the query domain. We also use the attributes along with the on-the-fly classifier to score the database images and obtain a hybrid ranking. We show interesting qualitative results and demonstrate by experiments with standard datasets that the proposed method improves upon the baseline on-the-fly classification system.

International audience

We propose a new model for recognizing human attributes (e.g. wearing a suit, sitting, short hair) and actions (e.g. running, riding a horse) in still images. The proposed model relies on a collection of part templates which are learnt discriminatively to explain specific scale-space locations in the images (in human centric coordinates). It avoids the limitations of highly structured models, which consist of a few (i.e. a mixture of) 'average' templates. To learn our model, we propose an algorithm which automatically mines out parts and learns corresponding discriminative templates with their respective locations from a large number of candidate parts. We validate the method on recent challenging datasets: (i) Willow 7 actions [7], (ii) 27 Human Attributes (HAT) [25], and (iii) Stanford 40 actions [37]. We obtain convincing qualitative and state-of-the-art quantitative results on the three datasets.

In many visual classification tasks the spatial distribution of discriminative information is (i) non uniform e.g. person `reading' can be distinguished from `taking a photo' based on the area around the arms i.e. ignoring the legs and (ii) has intra class variations e.g. different readers may hold the books differently. Motivated by these observations, we propose to learn the discriminative spatial saliency of images while simultaneously learning a max margin classifier for a given visual classification task. Using the saliency maps to weight the corresponding visual features improves the discriminative power of the image representation. We treat the saliency maps as latent variables and allow them to adapt to the image content to maximize the classification score, while regularizing the change in the saliency maps. Our experimental results on three challenging datasets, for (i) human action classification, (ii) fine grained classification and (iii) scene classification, demonstrate the effectiveness and wide applicability of the method.

International audience

We present a classifier unifying local features based representation and subspace based learning. We also propose a novel method to merge kernel eigen spaces (KES) in feature space. Subspace methods have traditionally been used with the full appearance of the image. Recently local features based bag-of-features (BoF) representation has performed impressively on classification tasks. We use KES with BoF vectors to construct class specific subspaces and use the distance of a query vector from the database KESs as the classification criteria. The use of local features makes our approach invariant to illumination, rotation, scale, small affine transformation and partial occlusions. The system allows hierarchy by merging the KES in the feature space. The classifier performs competitively on the challenging Caltech-101 dataset under normal and simulated occlusion conditions. We show hierarchy on a dataset of videos collected over the internet.