Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

As deep networks are applied to an ever-expanding set of computer vision tasks, protecting general privacy in image data has become a critically important goal. This paper presents a new framework for privacy-preserving data sharing that is robust to adversarial attacks and overcomes the known issues existing in previous approaches. We introduce the concept of a Deep Poisoning Function (DPF), which is a module inserted into a pre-trained deep network designed to perform a specific vision task. The DPF is optimized to deliberately poison image data to prevent known adversarial attacks, while ensuring that the altered image data is functionally equivalent to the non-poisoned data for the original task. Given this equivalence, both poisoned and non-poisoned data can be used for further retraining or fine-tuning. Experimental results on image classification and face recognition tasks prove the efficacy of the proposed method.

Crowd sourcing has become a widely adopted scheme to collect ground truth labels. However, it is a well-known problem that these labels can be very noisy. In this paper, we demonstrate how to learn a deep convolutional neural network (DCNN) from noisy labels, using facial expression recognition as an example. More specifically, we have 10 taggers to label each input image, and compare four different approaches to utilizing the multiple labels: majority voting, multi-label learning, probabilistic label drawing, and cross-entropy loss. We show that the traditional majority voting scheme does not perform as well as the last two approaches that fully leverage the label distribution. An enhanced FER+ data set with multiple labels for each face image will also be shared with the research community.

Acoustic events produced in meeting environments may contain useful information for perceptually aware interfaces and multimodal behavior analysis. In this paper, a system to detect and recognize these events from a multimodal perspective is presented combining information from multiple cameras and microphones. First, spectral and temporal features are extracted from a single audio channel and spatial localization is achieved by exploiting cross-correlation among microphone arrays. Second, several video cues obtained from multiperson tracking, motion analysis, face recognition, and object detection provide the visual counterpart of the acoustic events to be detected. A multimodal data fusion at score level is carried out using two approaches: weighted mean average and fuzzy integral. Finally, a multimodal database containing a rich variety of acoustic events has been recorded including manual annotations of the data. A set of metrics allow assessing the performance of the presented algorithms. This dataset is made publicly available for research purposes.

When applied to interactive seminars, the detection of acoustic events from only audio information shows a large amount of errors, which are mostly due to the temporal overlaps of sounds. Video signals may be a useful additional source of information to cope with that problem for particular events. In this work, we aim at improving the detection of steps by using two audio-based Acoustic Event Detection (AED) systems, with SVM and HMM, and a video-based AED system, which employs the output of a 3D video tracking algorithm. The fuzzy integral is used to fuse the outputs of the three detection systems. Experimental results using the CLEAR 2007 evaluation data show that video information can be successfully used to improve the results of audio-based AED.

This paper presents two approaches to the problem of simultaneous tracking of several people in low resolution sequences from multiple calibrated cameras. Spatial redundancy is exploited to generate a discrete 3D binary representation of the foreground objects in the scene. Color information obtained from a zenithal camera view is added to this 3D information. The first tracking approach implements heuristic association rules between blobs labelled according to spatiotemporal connectivity criteria. Association rules are based on a cost function which considers their placement and color histogram. In the second approach, a particle filtering scheme adapted to the incoming 3D discrete data is proposed. A volume likelihood function and a discrete 3D re-sampling procedure are introduced to evaluate and drive particles. Multiple targets are tracked by means of multiple particle filters and interaction among them is modeled through a 3D blocking scheme. Evaluation over the CLEAR 2007 database yields quantitative results assessing the performance of the proposed algorithm for indoor scenarios.

This paper presents a general analysis framework towards exploiting the underlying hierarchical and scalable structure of an articulated object for pose estimation and tracking. The Scalable Human Body Model (SHBM) is presented as a set of human body models ordered following a hierarchy criteria. The concept of annealing is applied to derive a generic particle filtering scheme able to perform a sequential filtering over the models contained in the SHBM leading to a structural annealing process. This scheme is applied to perform human motion capture in a multi-camera environment. Finally, the effectiveness of the proposed system is addressed by comparing its performance with the standard and annealed particle filtering approaches over an annotated database.

This paper presents an overview of our work on real-time multimodal tracking focus of attention of multiple persons in a SmartRoom scenario. Redundancy among cameras is exploited to generate a 3D discrete reconstruction of the space. This information is fed to a novel low complexity Monte Carlo based tracking scheme. Estimated locations of people in the room are used to automatically determine their head positions. Head orientation of every person is computed using video and audio separately and then a multimodal estimation is produced by combining data at feature level employing a decentralized Kalman filter. Finally, participantspsila focus attention is estimated by means of two geometric descriptors: the attention cone and the attention map. Experiments conducted over annotated databases yield quantitative results proving the effectiveness of the presented approach.

This article presents a multimodal approach to head pose estimation of individuals in environments equipped with multiple cameras and microphones, such as SmartRooms or automatic video conferencing. Determining the individuals head orientation is the basis for many forms of more sophisticated interactions between humans and technical devices and can also be used for automatic sensor selection (camera, microphone) in communications or video surveillance systems. The use of particle filters as a unified framework for the estimation of the head orientation for both monomodal and multimodal cases is proposed. In video, we estimate head orientation from color information by exploiting spatial redundancy among cameras. Audio information is processed to estimate the direction of the voice produced by a speaker making use of the directivity characteristics of the head radiation pattern. Furthermore, two different particle filter multimodal information fusion schemes for combining the audio and video streams are analyzed in terms of accuracy and robustness. In the first one, fusion is performed at a decision level by combining each monomodal head pose estimation, while the second one uses a joint estimation system combining information at data level. Experimental results conducted over the CLEAR 2006 evaluation database are reported and the comparison of the proposed multimodal head pose estimation algorithms with the reference monomodal approaches proves the effectiveness of the proposed approach.

Reliable measures of person positions are needed for computational perception of human activities taking place in a smart-room environment. In this work, we present the Person Tracking systems developed at UPC for audio, video and audio-video modalities in the context of the EU funded CHIL project research activities. The aim of the designed systems, and particularly of the new contributions proposed, is to deal robustly in both single and multiperson localization tasks independently on the environmental conditions. Besides the technology description, experimental results conducted for the CLEAR evaluation workshop are also reported.