Arnaud Dapogny

Edouard Yvinec, Arnaud Dapogny, Matthieu Cord et al.

Computationally expensive neural networks are ubiquitous in computer vision and solutions for efficient inference have drawn a growing attention in the machine learning community. Examples of such solutions comprise quantization, i.e. converting the processing values (weights and inputs) from floating point into integers e.g. int8 or int4. Concurrently, the rise of privacy concerns motivated the study of less invasive acceleration methods, such as data-free quantization of pre-trained models weights and activations. Previous approaches either exploit statistical information to deduce scalar ranges and scaling factors for the activations in a static manner, or dynamically adapt this range on-the-fly for each input of each layers (also referred to as activations): the latter generally being more accurate at the expanse of significantly slower inference. In this work, we argue that static input quantization can reach the accuracy levels of dynamic methods by means of a per-channel input quantization scheme that allows one to more finely preserve cross-channel dynamics. We show through a thorough empirical evaluation on multiple computer vision problems (e.g. ImageNet classification, Pascal VOC object detection as well as CityScapes semantic segmentation) that the proposed method, dubbed SPIQ, achieves accuracies rivalling dynamic approaches with static-level inference speed, significantly outperforming state-of-the-art quantization methods on every benchmark.

Edouard Yvinec, Arnaud Dapogny, Matthieu Cord et al.

Deep neural networks (DNNs) are nowadays ubiquitous in many domains such as computer vision. However, due to their high latency, the deployment of DNNs hinges on the development of compression techniques such as quantization which consists in lowering the number of bits used to encode the weights and activations. Growing concerns for privacy and security have motivated the development of data-free techniques, at the expanse of accuracy. In this paper, we identity the uniformity of the quantization operator as a limitation of existing approaches, and propose a data-free non-uniform method. More specifically, we argue that to be readily usable without dedicated hardware and implementation, non-uniform quantization shall not change the nature of the mathematical operations performed by the DNN. This leads to search among the continuous automorphisms of $(\mathbb{R}_+^*,\times)$, which boils down to the power functions defined by their exponent. To find this parameter, we propose to optimize the reconstruction error of each layer: in particular, we show that this procedure is locally convex and admits a unique solution. At inference time, we show that our approach, dubbed PowerQuant, only require simple modifications in the quantized DNN activation functions. As such, with only negligible overhead, it significantly outperforms existing methods in a variety of configurations.

Edouard Yvinec, Arnaud Dapogny, Matthieu Cord et al.

The leap in performance in state-of-the-art computer vision methods is attributed to the development of deep neural networks. However it often comes at a computational price which may hinder their deployment. To alleviate this limitation, structured pruning is a well known technique which consists in removing channels, neurons or filters, and is commonly applied in order to produce more compact models. In most cases, the computations to remove are selected based on a relative importance criterion. At the same time, the need for explainable predictive models has risen tremendously and motivated the development of robust attribution methods that highlight the relative importance of pixels of an input image or feature map. In this work, we discuss the limitations of existing pruning heuristics, among which magnitude and gradient-based methods. We draw inspiration from attribution methods to design a novel integrated gradient pruning criterion, in which the relevance of each neuron is defined as the integral of the gradient variation on a path towards this neuron removal. Furthermore, we propose an entwined DNN pruning and fine-tuning flowchart to better preserve DNN accuracy while removing parameters. We show through extensive validation on several datasets, architectures as well as pruning scenarios that the proposed method, dubbed SInGE, significantly outperforms existing state-of-the-art DNN pruning methods.

Gauthier Tallec, Edouard Yvinec, Arnaud Dapogny et al.

Action Unit (AU) Detection is the branch of affective computing that aims at recognizing unitary facial muscular movements. It is key to unlock unbiased computational face representations and has therefore aroused great interest in the past few years. One of the main obstacles toward building efficient deep learning based AU detection system is the lack of wide facial image databases annotated by AU experts. In that extent the ABAW challenge paves the way toward better AU detection as it involves a 2M frames AU annotated dataset. In this paper, we present our submission to the ABAW3 challenge. In a nutshell, we applied a multi-label detection transformer that leverage multi-head attention to learn which part of the face image is the most relevant to predict each AU.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly

Batch-Normalization (BN) layers have become fundamental components in the evermore complex deep neural network architectures. Such models require acceleration processes for deployment on edge devices. However, BN layers add computation bottlenecks due to the sequential operation processing: thus, a key, yet often overlooked component of the acceleration process is BN layers folding. In this paper, we demonstrate that the current BN folding approaches are suboptimal in terms of how many layers can be removed. We therefore provide a necessary and sufficient condition for BN folding and a corresponding optimal algorithm. The proposed approach systematically outperforms existing baselines and allows to dramatically reduce the inference time of deep neural networks.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly

Deep neural network (DNN) deployment has been confined to larger hardware devices due to their expensive computational requirements. This challenge has recently reached another scale with the emergence of large language models (LLMs). In order to reduce both their memory footprint and latency, a promising technique is quantization. It consists in converting floating point representations to low bit-width fixed point representations, usually by assuming a uniform mapping onto a regular grid. This process, referred to in the literature as uniform quantization, may however be ill-suited as most DNN weights and activations follow a bell-shaped distribution. This is even worse on LLMs whose weight distributions are known to exhibit large, high impact, outlier values. In this work, we propose an improvement over the most commonly adopted way to tackle this limitation in deep learning models quantization, namely, non-uniform quantization. NUPES leverages automorphisms to preserve the scalar multiplications. Such transformations are derived from power functions. However, the optimization of the exponent parameter and weight values remains a challenging and novel problem which could not be solved with previous post training optimization techniques which only learn to round up or down weight values in order to preserve the predictive function. We circumvent this limitation with a new paradigm: learning new quantized weights over the entire quantized space. Similarly, we enable the optimization of the power exponent, i.e. the optimization of the quantization operator itself during training by alleviating all the numerical instabilities. The resulting predictive function is compatible with integer-only low-bit inference. We show the ability of the method to achieve state-of-the-art compression rates in both, data-free and data-driven configurations.

Jules Bonnard, Arnaud Dapogny, Ferdinand Dhombres et al.

Facial Expression Recognition (FER) is crucial in many research domains because it enables machines to better understand human behaviours. FER methods face the problems of relatively small datasets and noisy data that don't allow classical networks to generalize well. To alleviate these issues, we guide the model to concentrate on specific facial areas like the eyes, the mouth or the eyebrows, which we argue are decisive to recognise facial expressions. We propose the Privileged Attribution Loss (PAL), a method that directs the attention of the model towards the most salient facial regions by encouraging its attribution maps to correspond to a heatmap formed by facial landmarks. Furthermore, we introduce several channel strategies that allow the model to have more degrees of freedom. The proposed method is independent of the backbone architecture and doesn't need additional semantic information at test time. Finally, experimental results show that the proposed PAL method outperforms current state-of-the-art methods on both RAF-DB and AffectNet.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly et al.

Deep neural networks (DNNs) demonstrate outstanding performance across most computer vision tasks. Some critical applications, such as autonomous driving or medical imaging, also require investigation into their behavior and the reasons behind the decisions they make. In this vein, DNN attribution consists in studying the relationship between the predictions of a DNN and its inputs. Attribution methods have been adapted to highlight the most relevant weights or neurons in a DNN, allowing to more efficiently select which weights or neurons can be pruned. However, a limitation of these approaches is that weights are typically compared within each layer separately, while some layers might appear as more critical than others. In this work, we propose to investigate DNN layer importance, i.e. to estimate the sensitivity of the accuracy w.r.t. perturbations applied at the layer level. To do so, we propose a novel dataset to evaluate our method as well as future works. We benchmark a number of criteria and draw conclusions regarding how to assess DNN layer importance and, consequently, how to budgetize layers for increased DNN efficiency (with applications for DNN pruning and quantization), as well as robustness to hardware failure (e.g. bit swaps).

Gauthier Tallec, Edouard Yvinec, Arnaud Dapogny et al.

The rising performance of deep neural networks is often empirically attributed to an increase in the available computational power, which allows complex models to be trained upon large amounts of annotated data. However, increased model complexity leads to costly deployment of modern neural networks, while gathering such amounts of data requires huge costs to avoid label noise. In this work, we study the ability of compression methods to tackle both of these problems at once. We hypothesize that quantization-aware training, by restricting the expressivity of neural networks, behaves as a regularization. Thus, it may help fighting overfitting on noisy data while also allowing for the compression of the model at inference. We first validate this claim on a controlled test with manually introduced label noise. Furthermore, we also test the proposed method on Facial Action Unit detection, where labels are typically noisy due to the subtlety of the task. In all cases, our results suggests that quantization significantly improve the results compared with existing baselines, regularization as well as other compression methods.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly

The massive interest in deep neural networks (DNNs) for both computer vision and natural language processing has been sparked by the growth in computational power. However, this led to an increase in the memory footprint, to a point where it can be challenging to simply load a model on commodity devices such as mobile phones. To address this limitation, quantization is a favored solution as it maps high precision tensors to a low precision, memory efficient format. In terms of memory footprint reduction, its most effective variants are based on codebooks. These methods, however, suffer from two limitations. First, they either define a single codebook for each tensor, or use a memory-expensive mapping to multiple codebooks. Second, gradient descent optimization of the mapping favors jumps toward extreme values, hence not defining a proximal search. In this work, we propose to address these two limitations. First, we initially group similarly distributed neurons and leverage the re-ordered structure to either apply different scale factors to the different groups, or map weights that fall in these groups to several codebooks, without any mapping overhead. Second, stemming from this initialization, we propose a joint learning of the codebook and weight mappings that bears similarities with recent gradient-based post-training quantization techniques. Third, drawing estimation from straight-through estimation techniques, we introduce a novel gradient update definition to enable a proximal search of the codebooks and their mappings. The proposed jointly learnable codebooks and mappings (JLCM) method allows a very efficient approximation of any DNN: as such, a Llama 7B can be compressed down to 2Go and loaded on 5-year-old smartphones.

Eden Belouadah, Arnaud Dapogny, Kevin Bailly

Class-Incremental learning (CIL) refers to the ability of artificial agents to integrate new classes as they appear in a stream. It is particularly interesting in evolving environments where agents have limited access to memory and computational resources. The main challenge of incremental learning is catastrophic forgetting, the inability of neural networks to retain past knowledge when learning a new one. Unfortunately, most existing class-incremental methods for object detection are applied to two-stage algorithms such as Faster-RCNN, and rely on rehearsal memory to retain past knowledge. We argue that those are not suitable in resource-limited environments, and more effort should be dedicated to anchor-free and rehearsal-free object detection. In this paper, we propose MultIOD, a class-incremental object detector based on CenterNet. Our contributions are: (1) we propose a multihead feature pyramid and multihead detection architecture to efficiently separate class representations, (2) we employ transfer learning between classes learned initially and those learned incrementally to tackle catastrophic forgetting, and (3) we use a class-wise non-max-suppression as a post-processing technique to remove redundant boxes. Results show that our method outperforms state-of-the-art methods on two Pascal VOC datasets, while only saving the model in its current state, contrary to other distillation-based counterparts.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly

Quantization has become a crucial step for the efficient deployment of deep neural networks, where floating point operations are converted to simpler fixed point operations. In its most naive form, it simply consists in a combination of scaling and rounding transformations, leading to either a limited compression rate or a significant accuracy drop. Recently, Gradient-based post-training quantization (GPTQ) methods appears to be constitute a suitable trade-off between such simple methods and more powerful, yet expensive Quantization-Aware Training (QAT) approaches, particularly when attempting to quantize LLMs, where scalability of the quantization process is of paramount importance. GPTQ essentially consists in learning the rounding operation using a small calibration set. In this work, we challenge common choices in GPTQ methods. In particular, we show that the process is, to a certain extent, robust to a number of variables (weight selection, feature augmentation, choice of calibration set). More importantly, we derive a number of best practices for designing more efficient and scalable GPTQ methods, regarding the problem formulation (loss, degrees of freedom, use of non-uniform quantization schemes) or optimization process (choice of variable and optimizer). Lastly, we propose a novel importance-based mixed-precision technique. Those guidelines lead to significant performance improvements on all the tested state-of-the-art GPTQ methods and networks (e.g. +6.819 points on ViT for 4-bit quantization), paving the way for the design of scalable, yet effective quantization methods.

Gauthier Tallec, Arnaud Dapogny, Kevin Bailly

Action Unit (AU) detection aims at automatically caracterizing facial expressions with the muscular activations they involve. Its main interest is to provide a low-level face representation that can be used to assist higher level affective computing tasks learning. Yet, it is a challenging task. Indeed, the available databases display limited face variability and are imbalanced toward neutral expressions. Furthermore, as AU involve subtle face movements they are difficult to annotate so that some of the few provided datapoints may be mislabeled. In this work, we aim at exploiting label smoothing ability to mitigate noisy examples impact by reducing confidence [1]. However, applying label smoothing as it is may aggravate imbalance-based pre-existing under-confidence issue and degrade performance. To circumvent this issue, we propose Robin Hood Label Smoothing (RHLS). RHLS principle is to restrain label smoothing confidence reduction to the majority class. In that extent, it alleviates both the imbalance-based over-confidence issue and the negative impact of noisy majority class examples. From an experimental standpoint, we show that RHLS provides a free performance improvement in AU detection. In particular, by applying it on top of a modern multi-task baseline we get promising results on BP4D and outperform state-of-the-art methods on DISFA.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly

Deep neural networks (DNNs) are ubiquitous in computer vision and natural language processing, but suffer from high inference cost. This problem can be addressed by quantization, which consists in converting floating point perations into a lower bit-width format. With the growing concerns on privacy rights, we focus our efforts on data-free methods. However, such techniques suffer from their lack of adaptability to the target devices, as a hardware typically only support specific bit widths. Thus, to adapt to a variety of devices, a quantization method shall be flexible enough to find good accuracy v.s. speed trade-offs for every bit width and target device. To achieve this, we propose PIPE, a quantization method that leverages residual error expansion, along with group sparsity and an ensemble approximation for better parallelization. PIPE is backed off by strong theoretical guarantees and achieves superior performance on every benchmarked application (from vision to NLP tasks), architecture (ConvNets, transformers) and bit-width (from int8 to ternary quantization).

Rémi Ouazan Reboul, Edouard Yvinec, Arnaud Dapogny et al.

Deep neural networks (DNNs) have become ubiquitous in addressing a number of problems, particularly in computer vision. However, DNN inference is computationally intensive, which can be prohibitive e.g. when considering edge devices. To solve this problem, a popular solution is DNN pruning, and more so structured pruning, where coherent computational blocks (e.g. channels for convolutional networks) are removed: as an exhaustive search of the space of pruned sub-models is intractable in practice, channels are typically removed iteratively based on an importance estimation heuristic. Recently, promising latency-aware pruning methods were proposed, where channels are removed until the network reaches a target budget of wall-clock latency pre-emptively estimated on specific hardware. In this paper, we present Archtree, a novel method for latency-driven structured pruning of DNNs. Archtree explores multiple candidate pruned sub-models in parallel in a tree-like fashion, allowing for a better exploration of the search space. Furthermore, it involves on-the-fly latency estimation on the target hardware, accounting for closer latencies as compared to the specified budget. Empirical results on several DNN architectures and target hardware show that Archtree better preserves the original model accuracy while better fitting the latency budget as compared to existing state-of-the-art methods.

Edouard Yvinec, Arnaud Dapogny, Kevin Bailly

Deep neural networks (DNNs) offer the highest performance in a wide range of applications in computer vision. These results rely on over-parameterized backbones, which are expensive to run. This computational burden can be dramatically reduced by quantizing (in either data-free (DFQ), post-training (PTQ) or quantization-aware training (QAT) scenarios) floating point values to ternary values (2 bits, with each weight taking value in {-1,0,1}). In this context, we observe that rounding to nearest minimizes the expected error given a uniform distribution and thus does not account for the skewness and kurtosis of the weight distribution, which strongly affects ternary quantization performance. This raises the following question: shall one minimize the highest or average quantization error? To answer this, we design two operators: TQuant and MQuant that correspond to these respective minimization tasks. We show experimentally that our approach allows to significantly improve the performance of ternary quantization through a variety of scenarios in DFQ, PTQ and QAT and give strong insights to pave the way for future research in deep neural network quantization.

Gauthier Tallec, Jules Bonnard, Arnaud Dapogny et al.

Face based affective computing consists in detecting emotions from face images. It is useful to unlock better automatic comprehension of human behaviours and could pave the way toward improved human-machines interactions. However it comes with the challenging task of designing a computational representation of emotions. So far, emotions have been represented either continuously in the 2D Valence/Arousal space or in a discrete manner with Ekman's 7 basic emotions. Alternatively, Ekman's Facial Action Unit (AU) system have also been used to caracterize emotions using a codebook of unitary muscular activations. ABAW3 and ABAW4 Multi-Task Challenges are the first work to provide a large scale database annotated with those three types of labels. In this paper we present a transformer based multi-task method for jointly learning to predict valence arousal, action units and basic emotions. From an architectural standpoint our method uses a taskwise token approach to efficiently model the similarities between the tasks. From a learning point of view we use an uncertainty weighted loss for modelling the difference of stochasticity between the three tasks annotations.

Hugo Caselles-Dupré, Mathis Koroglu, Guillaume Jeanneret et al.

Diffusion-based image-to-video (I2V) models are increasingly effective, yet they struggle to scale to ultra-high-resolution inputs (e.g., 4K). Generating videos at the model's native resolution often loses fine-grained structure, whereas high-resolution tiled denoising preserves local detail but breaks global layout consistency. This failure mode is particularly severe in the fresco animation setting: monumental artworks containing many distinct characters, objects, and semantically different sub-scenes that must remain spatially coherent over time. We introduce FrescoDiffusion, a training-free method for coherent large-format I2V generation from a single complex image. The key idea is to augment tiled denoising with a precomputed latent prior: we first generate a low-resolution video at the underlying model resolution and upsample its latent trajectory to obtain a global reference that captures long-range temporal and spatial structure. For 4K generation, we compute per-tile noise predictions and fuse them with this reference at every diffusion timestep by minimizing a single weighted least-squares objective in model-output space. The objective combines a standard tile-merging criterion with our regularization term, yielding a closed-form fusion update that strengthens global coherence while retaining fine detail. We additionally provide a spatial regularization variable that enables region-level control over where motion is allowed. Experiments on the VBench-I2V dataset and our proposed fresco I2V dataset show improved global consistency and fidelity over tiled baselines, while being computationally efficient. Our regularization enables explicit controllability of the trade-off between creativity and consistency.

Jayneel Parekh, Pegah Khayatan, Mustafa Shukor et al.

Steering has emerged as a practical approach to enable post-hoc guidance of LLMs towards enforcing a specific behavior. However, it remains largely underexplored for multimodal LLMs (MLLMs); furthermore, existing steering techniques, such as mean steering, rely on a single steering vector, applied independently of the input query. This paradigm faces limitations when the desired behavior is dependent on the example at hand. For example, a safe answer may consist in abstaining from answering when asked for an illegal activity, or may point to external resources or consultation with an expert when asked about medical advice. In this paper, we investigate a fine-grained steering that uses an input-specific linear shift. This shift is computed using contrastive input-specific prompting. However, the input-specific prompts required for this approach are not known at test time. Therefore, we propose to train a small auxiliary module to predict the input-specific steering vector. Our approach, dubbed as L2S (Learn-to-Steer), demonstrates that it reduces hallucinations and enforces safety in MLLMs, outperforming other static baselines. Our code is publicly available at https://jayneelparekh.github.io/learn-to-steer/

Pegah Khayatan, Jayneel Parekh, Arnaud Dapogny et al.

Despite impressive progress in capabilities of large vision-language models (LVLMs), these systems remain vulnerable to hallucinations, i.e., outputs that are not grounded in the visual input. Prior work has attributed hallucinations in LVLMs to factors such as limitations of the vision backbone or the dominance of the language component, yet the relative importance of these factors remains unclear. To resolve this ambiguity, We propose HalluScope, a benchmark to better understand the extent to which different factors induce hallucinations. Our analysis indicates that hallucinations largely stem from excessive reliance on textual priors and background knowledge, especially information introduced through textual instructions. To mitigate hallucinations induced by textual instruction priors, we propose HalluVL-DPO, a framework for fine-tuning off-the-shelf LVLMs towards more visually grounded responses. HalluVL-DPO leverages preference optimization using a curated training dataset that we construct, guiding the model to prefer grounded responses over hallucinated ones. We demonstrate that our optimized model effectively mitigates the targeted hallucination failure mode, while preserving or improving performance on other hallucination benchmarks and visual capability evaluations. To support reproducibility and further research, we will publicly release our evaluation benchmark, preference training dataset, and code at https://pegah-kh.github.io/projects/prompts-override-vision/ .

Pegah Khayatan, Mustafa Shukor, Jayneel Parekh et al.

Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.

Gauthier Tallec, Arnaud Dapogny, Kevin Bailly

Action Units (AU) are muscular activations used to describe facial expressions. Therefore accurate AU recognition unlocks unbiaised face representation which can improve face-based affective computing applications. From a learning standpoint AU detection is a multi-task problem with strong inter-task dependencies. To solve such problem, most approaches either rely on weight sharing, or add explicit dependency modelling by decomposing the joint task distribution using Bayes chain rule. If the latter strategy yields comprehensive inter-task relationships modelling, it requires imposing an arbitrary order into an unordered task set. Crucially, this ordering choice has been identified as a source of performance variations. In this paper, we present Multi-Order Network (MONET), a multi-task method with joint task order optimization. MONET uses a differentiable order selection to jointly learn task-wise modules with their optimal chaining order. Furthermore, we introduce warmup and order dropout to enhance order selection by encouraging order exploration. Experimentally, we first demonstrate MONET capacity to retrieve the optimal order in a toy environment. Second, we validate MONET architecture by showing that MONET outperforms existing multi-task baselines on multiple attribute detection problems chosen for their wide range of dependency settings. More importantly, we demonstrate that MONET significantly extends state-of-the-art performance in AU detection.

Edouard Yvinec, Arnaud Dapogny, Matthieu Cord et al.

Pruning Deep Neural Networks (DNNs) is a prominent field of study in the goal of inference runtime acceleration. In this paper, we introduce a novel data-free pruning protocol RED++. Only requiring a trained neural network, and not specific to DNN architecture, we exploit an adaptive data-free scalar hashing which exhibits redundancies among neuron weight values. We study the theoretical and empirical guarantees on the preservation of the accuracy from the hashing as well as the expected pruning ratio resulting from the exploitation of said redundancies. We propose a novel data-free pruning technique of DNN layers which removes the input-wise redundant operations. This algorithm is straightforward, parallelizable and offers novel perspective on DNN pruning by shifting the burden of large computation to efficient memory access and allocation. We provide theoretical guarantees on RED++ performance and empirically demonstrate its superiority over other data-free pruning methods and its competitiveness with data-driven ones on ResNets, MobileNets and EfficientNets.

Arthur Douillard, Yifu Chen, Arnaud Dapogny et al.

Deep learning approaches are nowadays ubiquitously used to tackle computer vision tasks such as semantic segmentation, requiring large datasets and substantial computational power. Continual learning for semantic segmentation (CSS) is an emerging trend that consists in updating an old model by sequentially adding new classes. However, continual learning methods are usually prone to catastrophic forgetting. This issue is further aggravated in CSS where, at each step, old classes from previous iterations are collapsed into the background. In this paper, we propose Local POD, a multi-scale pooling distillation scheme that preserves long- and short-range spatial relationships at feature level. Furthermore, we design an entropy-based pseudo-labelling of the background w.r.t. classes predicted by the old model to deal with background shift and avoid catastrophic forgetting of the old classes. Finally, we introduce a novel rehearsal method that is particularly suited for segmentation. Our approach, called PLOP, significantly outperforms state-of-the-art methods in existing CSS scenarios, as well as in newly proposed challenging benchmarks.

Edouard Yvinec, Arnaud Dapogny, Matthieu Cord et al.

Deep Neural Networks (DNNs) are ubiquitous in today's computer vision land-scape, despite involving considerable computational costs. The mainstream approaches for runtime acceleration consist in pruning connections (unstructured pruning) or, better, filters (structured pruning), both often requiring data to re-train the model. In this paper, we present RED, a data-free structured, unified approach to tackle structured pruning. First, we propose a novel adaptive hashing of the scalar DNN weight distribution densities to increase the number of identical neurons represented by their weight vectors. Second, we prune the network by merging redundant neurons based on their relative similarities, as defined by their distance. Third, we propose a novel uneven depthwise separation technique to further prune convolutional layers. We demonstrate through a large variety of benchmarks that RED largely outperforms other data-free pruning methods, often reaching performance similar to unconstrained, data-driven methods.

Arthur Douillard, Yifu Chen, Arnaud Dapogny et al.

Deep learning approaches are nowadays ubiquitously used to tackle computer vision tasks such as semantic segmentation, requiring large datasets and substantial computational power. Continual learning for semantic segmentation (CSS) is an emerging trend that consists in updating an old model by sequentially adding new classes. However, continual learning methods are usually prone to catastrophic forgetting. This issue is further aggravated in CSS where, at each step, old classes from previous iterations are collapsed into the background. In this paper, we propose Local POD, a multi-scale pooling distillation scheme that preserves long- and short-range spatial relationships at feature level. Furthermore, we design an entropy-based pseudo-labelling of the background w.r.t. classes predicted by the old model to deal with background shift and avoid catastrophic forgetting of the old classes. Our approach, called PLOP, significantly outperforms state-of-the-art methods in existing CSS scenarios, as well as in newly proposed challenging benchmarks.

Estephe Arnaud, Arnaud Dapogny, Kevin Bailly

For a number of machine learning problems, an exogenous variable can be identified such that it heavily influences the appearance of the different classes, and an ideal classifier should be invariant to this variable. An example of such exogenous variable is identity if facial expression recognition (FER) is considered. In this paper, we propose a dual exogenous/endogenous representation. The former captures the exogenous variable whereas the second one models the task at hand (e.g. facial expression). We design a prediction layer that uses a tree-gated deep ensemble conditioned by the exogenous representation. We also propose an exogenous dispelling loss to remove the exogenous information from the endogenous representation. Thus, the exogenous information is used two times in a throwable fashion, first as a conditioning variable for the target task, and second to create invariance within the endogenous representation. We call this method THIN, standing for THrowable Information Networks. We experimentally validate THIN in several contexts where an exogenous information can be identified, such as digit recognition under large rotations and shape recognition at multiple scales. We also apply it to FER with identity as the exogenous variable. We demonstrate that THIN significantly outperforms state-of-the-art approaches on several challenging datasets.

Edouard Yvinec, Arnaud Dapogny, Kévin Bailly

From medical research to gaming applications, gaze estimation is becoming a valuable tool. While there exists a number of hardware-based solutions, recent deep learning-based approaches, coupled with the availability of large-scale databases, have allowed to provide a precise gaze estimate using only consumer sensors. However, there remains a number of questions, regarding the problem formulation, architectural choices and learning paradigms for designing gaze estimation systems in order to bridge the gap between geometry-based systems involving specific hardware and approaches using consumer sensors only. In this paper, we introduce a deep, end-to-end trainable ensemble of heatmap-based weak predictors for 2D/3D gaze estimation. We show that, through heterogeneous architectural design of these weak predictors, we can improve the decorrelation between the latter predictors to design more robust deep ensemble models. Furthermore, we propose a stochastic combinatory loss that consists in randomly sampling combinations of weak predictors at train time. This allows to train better individual weak predictors, with lower correlation between them. This, in turns, allows to significantly enhance the performance of the deep ensemble. We show that our Deep heterogeneous ensemble with Stochastic Combinatory loss (DeeSCo) outperforms state-of-the-art approaches for 2D/3D gaze estimation on multiple datasets.

Arnaud Dapogny, Kévin Bailly, Matthieu Cord

Head pose estimation and face alignment constitute a backbone preprocessing for many applications relying on face analysis. While both are closely related tasks, they are generally addressed separately, e.g. by deducing the head pose from the landmark locations. In this paper, we propose to entwine face alignment and head pose tasks inside an attentional cascade. This cascade uses a geometry transfer network for integrating heterogeneous annotations to enhance landmark localization accuracy. Furthermore, we propose a doubly-conditional fusion scheme to select relevant feature maps, and regions thereof, based on a current head pose and landmark localization estimate. We empirically show the benefit of entwining head pose and landmark localization objectives inside our architecture, and that the proposed AC-DC model enhances the state-of-the-art accuracy on multiple databases for both face alignment and head pose estimation tasks.

Estephe Arnaud, Arnaud Dapogny, Kevin Bailly

Face alignment consists of aligning a shape model on a face image. It is an active domain in computer vision as it is a preprocessing for a number of face analysis and synthesis applications. Current state-of-the-art methods already perform well on "easy" datasets, with moderate head pose variations, but may not be robust for "in-the-wild" data with poses up to 90°. In order to increase robustness to an ensemble of factors of variations (e.g. head pose or occlusions), a given layer (e.g. a regressor or an upstream CNN layer) can be replaced by a Mixture of Experts (MoE) layer that uses an ensemble of experts instead of a single one. The weights of this mixture can be learned as gating functions to jointly learn the experts and the corresponding weights. In this paper, we propose to use tree-structured gates which allows a hierarchical weighting of the experts (Tree-MoE). We investigate the use of Tree-MoE layers in different contexts in the frame of face alignment with cascaded regression, firstly for emphasizing relevant, more specialized feature extractors depending of a high-level semantic information such as head pose (Pose-Tree-MoE), and secondly as an overall more robust regression layer. We perform extensive experiments on several challenging face alignment datasets, demonstrating that our approach outperforms the state-of-the-art methods.

Estephe Arnaud, Arnaud Dapogny, Kevin Bailly

Face alignment consists in aligning a shape model on a face in an image. It is an active domain in computer vision as it is a preprocessing for applications like facial expression recognition, face recognition and tracking, face animation, etc. Current state-of-the-art methods already perform well on "easy" datasets, i.e. those that present moderate variations in head pose, expression, illumination or partial occlusions, but may not be robust to "in-the-wild" data. In this paper, we address this problem by using an ensemble of deep regressors instead of a single large regressor. Furthermore, instead of averaging the outputs of each regressor, we propose an adaptive weighting scheme that uses a tree-structured gate. Experiments on several challenging face datasets demonstrate that our approach outperforms the state-of-the-art methods.

Yifu Chen, Arnaud Dapogny, Matthieu Cord

While nowadays deep neural networks achieve impressive performances on semantic segmentation tasks, they are usually trained by optimizing pixel-wise losses such as cross-entropy. As a result, the predictions outputted by such networks usually struggle to accurately capture the object boundaries and exhibit holes inside the objects. In this paper, we propose a novel approach to improve the structure of the predicted segmentation masks. We introduce a novel semantic edge detection network, which allows to match the predicted and ground truth segmentation masks. This Semantic Edge-Aware strategy (SEMEDA) can be combined with any backbone deep network in an end-to-end training framework. Through thorough experimental validation on Pascal VOC 2012 and Cityscapes datasets, we show that the proposed SEMEDA approach enhances the structure of the predicted segmentation masks by enforcing sharp boundaries and avoiding discontinuities inside objects, improving the segmentation performance. In addition, our semantic edge-aware loss can be integrated into any popular segmentation network without requiring any additional annotation and with negligible computational load, as compared to standard pixel-wise cross-entropy loss.

Arnaud Dapogny, Matthieu Cord, Patrick Perez

Image completion is the problem of generating whole images from fragments only. It encompasses inpainting (generating a patch given its surrounding), reverse inpainting/extrapolation (generating the periphery given the central patch) as well as colorization (generating one or several channels given other ones). In this paper, we employ a deep network to perform image completion, with adversarial training as well as perceptual and completion losses, and call it the ``missing data encoder'' (MDE). We consider several configurations based on how the seed fragments are chosen. We show that training MDE for ``random extrapolation and colorization'' (MDE-REC), i.e. using random channel-independent fragments, allows a better capture of the image semantics and geometry. MDE training makes use of a novel ``hide-and-seek'' adversarial loss, where the discriminator seeks the original non-masked regions, while the generator tries to hide them. We validate our models both qualitatively and quantitatively on several datasets, showing their interest for image completion, unsupervised representation learning as well as face occlusion handling.

Arnaud Dapogny, Kévin Bailly, Matthieu Cord

Face Alignment is an active computer vision domain, that consists in localizing a number of facial landmarks that vary across datasets. State-of-the-art face alignment methods either consist in end-to-end regression, or in refining the shape in a cascaded manner, starting from an initial guess. In this paper, we introduce DeCaFA, an end-to-end deep convolutional cascade architecture for face alignment. DeCaFA uses fully-convolutional stages to keep full spatial resolution throughout the cascade. Between each cascade stage, DeCaFA uses multiple chained transfer layers with spatial softmax to produce landmark-wise attention maps for each of several landmark alignment tasks. Weighted intermediate supervision, as well as efficient feature fusion between the stages allow to learn to progressively refine the attention maps in an end-to-end manner. We show experimentally that DeCaFA significantly outperforms existing approaches on 300W, CelebA and WFLW databases. In addition, we show that DeCaFA can learn fine alignment with reasonable accuracy from very few images using coarsely annotated data.

Arnaud Dapogny, Kévin Bailly, Séverine Dubuisson

Face alignment is an active topic in computer vision, consisting in aligning a shape model on the face. To this end, most modern approaches refine the shape in a cascaded manner, starting from an initial guess. Those shape updates can either be applied in the feature point space (\textit{i.e.} explicit updates) or in a low-dimensional, parametric space. In this paper, we propose a semi-parametric cascade that first aligns a parametric shape, then captures more fine-grained deformations of an explicit shape. For the purpose of learning shape updates at each cascade stage, we introduce a deep greedy neural forest (GNF) model, which is an improved version of deep neural forest (NF). GNF appears as an ideal regressor for face alignment, as it combines differentiability, high expressivity and fast evaluation runtime. The proposed framework is very fast and achieves high accuracies on multiple challenging benchmarks, including small, medium and large pose experiments.

Arnaud Dapogny, Kévin Bailly, Séverine Dubuisson

Fully-Automatic Facial Expression Recognition (FER) from still images is a challenging task as it involves handling large interpersonal morphological differences, and as partial occlusions can occasionally happen. Furthermore, labelling expressions is a time-consuming process that is prone to subjectivity, thus the variability may not be fully covered by the training data. In this work, we propose to train Random Forests upon spatially defined local subspaces of the face. The output local predictions form a categorical expression-driven high-level representation that we call Local Expression Predictions (LEPs). LEPs can be combined to describe categorical facial expressions as well as Action Units (AUs). Furthermore, LEPs can be weighted by confidence scores provided by an autoencoder network. Such network is trained to locally capture the manifold of the non-occluded training data in a hierarchical way. Extensive experiments show that the proposed LEP representation yields high descriptive power for categorical expressions and AU occurrence prediction, and leads to interesting perspectives towards the design of occlusion-robust and confidence-aware FER systems.

Arnaud Dapogny, Kévin Bailly, Séverine Dubuisson

Automatic facial expression classification (FER) from videos is a critical problem for the development of intelligent human-computer interaction systems. Still, it is a challenging problem that involves capturing high-dimensional spatio-temporal patterns describing the variation of one's appearance over time. Such representation undergoes great variability of the facial morphology and environmental factors as well as head pose variations. In this paper, we use Conditional Random Forests to capture low-level expression transition patterns. More specifically, heterogeneous derivative features (e.g. feature point movements or texture variations) are evaluated upon pairs of images. When testing on a video frame, pairs are created between this current frame and previous ones and predictions for each previous frame are used to draw trees from Pairwise Conditional Random Forests (PCRF) whose pairwise outputs are averaged over time to produce robust estimates. Moreover, PCRF collections can also be conditioned on head pose estimation for multi-view dynamic FER. As such, our approach appears as a natural extension of Random Forests for learning spatio-temporal patterns, potentially from multiple viewpoints. Experiments on popular datasets show that our method leads to significant improvements over standard Random Forests as well as state-of-the-art approaches on several scenarios, including a novel multi-view video corpus generated from a publicly available database.