Weakly Supervised Object Localization (WSOL) allows for training deep learning models for classification and localization, using only global class-level labels. The lack of bounding box (bbox) supervision during training represents a considerable challenge for hyper-parameter search and model selection. Earlier WSOL works implicitly observed localization performance over a test set which leads to biased performance evaluation. More recently, a better WSOL protocol has been proposed, where a validation set with bbox annotations is held out for model selection. Although it does not rely on the test set, this protocol is unrealistic since bboxes are not available in real-world applications, and when available, it is better to use them directly to fit model weights. Our initial empirical analysis shows that the localization performance of a model declines significantly when using only image-class labels for model selection (compared to using bounding-box annotations). This suggests that adding bounding-box labels is preferable for selecting the best model for localization. In this paper, we introduce a new WSOL validation protocol that provides a localization signal without the need for manual bbox annotations. In particular, we leverage noisy pseudo boxes from an off-the-shelf ROI proposal generator such as Selective-Search, CLIP, and RPN pretrained models for model selection. Our experimental results with several WSOL methods on ILSVRC and CUB-200-2011 datasets show that our noisy boxes allow selecting models with performance close to those selected using ground truth boxes, and better than models selected using only image-class labels.
A common practice in deep learning consists of training large neural networks on massive datasets to perform accurately for different domains and tasks. While this methodology may work well in numerous application areas, it only applies across modalities due to a larger distribution shift in data captured using different sensors. This paper focuses on the problem of adapting a large object detection model to one or multiple modalities while being efficient. To do so, we propose ModTr as an alternative to the common approach of fine-tuning large models. ModTr consists of adapting the input with a small transformation network trained to minimize the detection loss directly. The original model can therefore work on the translated inputs without any further change or fine-tuning to its parameters. Experimental results on translating from IR to RGB images on two well-known datasets show that this simple ModTr approach provides detectors that can perform comparably or better than the standard fine-tuning without forgetting the original knowledge. This opens the doors to a more flexible and efficient service-based detection pipeline in which, instead of using a different detector for each modality, a unique and unaltered server is constantly running, where multiple modalities with the corresponding translations can query it. Code: https://github.com/heitorrapela/ModTr.
A key challenge in visible-infrared person re-identification (V-I ReID) is training a backbone model capable of effectively addressing the significant discrepancies across modalities. State-of-the-art methods that generate a single intermediate bridging domain are often less effective, as this generated domain may not adequately capture sufficient common discriminant information. This paper introduces the Bidirectional Multi-step Domain Generalization (BMDG), a novel approach for unifying feature representations across diverse modalities. BMDG creates multiple virtual intermediate domains by finding and aligning body part features extracted from both I and V modalities. Indeed, BMDG aims to reduce the modality gaps in two steps. First, it aligns modalities in feature space by learning shared and modality-invariant body part prototypes from V and I images. Then, it generalizes the feature representation by applying bidirectional multi-step learning, which progressively refines feature representations in each step and incorporates more prototypes from both modalities. In particular, our method minimizes the cross-modal gap by identifying and aligning shared prototypes that capture key discriminative features across modalities, then uses multiple bridging steps based on this information to enhance the feature representation. Experiments conducted on challenging V-I ReID datasets indicate that our BMDG approach outperforms state-of-the-art part-based models or methods that generate an intermediate domain from V-I person ReID.
Audiovisual emotion recognition (ER) in videos has immense potential over unimodal performance. It effectively leverages the inter- and intra-modal dependencies between visual and auditory modalities. This work proposes a novel audio-visual emotion recognition system utilizing a joint multimodal transformer architecture with key-based cross-attention. This framework aims to exploit the complementary nature of audio and visual cues (facial expressions and vocal patterns) in videos, leading to superior performance compared to solely relying on a single modality. The proposed model leverages separate backbones for capturing intra-modal temporal dependencies within each modality (audio and visual). Subsequently, a joint multimodal transformer architecture integrates the individual modality embeddings, enabling the model to effectively capture inter-modal (between audio and visual) and intra-modal (within each modality) relationships. Extensive evaluations on the challenging Affwild2 dataset demonstrate that the proposed model significantly outperforms baseline and state-of-the-art methods in ER tasks.
Domain adaptation methods for object detection (OD) strive to mitigate the impact of distribution shifts by promoting feature alignment across source and target domains. Multi-source domain adaptation (MSDA) allows leveraging multiple annotated source datasets, and unlabeled target data to improve the accuracy and robustness of the detection model. Most state-of-the-art MSDA methods for OD perform feature alignment in a class-agnostic manner. This is challenging since the objects have unique modal information due to variations in object appearance across domains. A recent prototype-based approach proposed a class-wise alignment, yet it suffers from error accumulation due to noisy pseudo-labels which can negatively affect adaptation with imbalanced data. To overcome these limitations, we propose an attention-based class-conditioned alignment scheme for MSDA that aligns instances of each object category across domains. In particular, an attention module coupled with an adversarial domain classifier allows learning domain-invariant and class-specific instance representations. Experimental results on multiple benchmarking MSDA datasets indicate that our method outperforms the state-of-the-art methods and is robust to class imbalance. Our code is available at https://github.com/imatif17/ACIA.
While state-of-the-art facial expression recognition (FER) classifiers achieve a high level of accuracy, they lack interpretability, an important aspect for end-users. To recognize basic facial expressions, experts resort to a codebook associating a set of spatial action units to a facial expression. In this paper, we follow the same expert footsteps, and propose a learning strategy that allows us to explicitly incorporate spatial action units (aus) cues into the classifier's training to build a deep interpretable model. In particular, using this aus codebook, input image expression label, and facial landmarks, a single action units heatmap is built to indicate the most discriminative regions of interest in the image w.r.t the facial expression. We leverage this valuable spatial cue to train a deep interpretable classifier for FER. This is achieved by constraining the spatial layer features of a classifier to be correlated with \aus map. Using a composite loss, the classifier is trained to correctly classify an image while yielding interpretable visual layer-wise attention correlated with aus maps, simulating the experts' decision process. This is achieved using only the image class expression as supervision and without any extra manual annotations. Moreover, our method is generic. It can be applied to any CNN- or transformer-based deep classifier without the need for architectural change or adding significant training time. Our extensive evaluation on two public benchmarks RAFDB, and AFFECTNET datasets shows that our proposed strategy can improve layer-wise interpretability without degrading classification performance. In addition, we explore a common type of interpretable classifiers that rely on Class-Activation Mapping methods (CAMs), and we show that our training technique improves the CAM interpretability.
Multimodal affect recognition models have reached remarkable performance in the lab environment due to their ability to model complementary and redundant semantic information. However, these models struggle in the wild, mainly because of the unavailability or quality of modalities used for training. In practice, only a subset of the training-time modalities may be available at test time. Learning with privileged information (PI) enables deep learning models (DL) to exploit data from additional modalities only available during training. State-of-the-art knowledge distillation (KD) methods have been proposed to distill multiple teacher models (each trained on a modality) to a common student model. These privileged KD methods typically utilize point-to-point matching and have no explicit mechanism to capture the structural information in the teacher representation space formed by introducing the privileged modality. We argue that encoding this same structure in the student space may lead to enhanced student performance. This paper introduces a new structural KD mechanism based on optimal transport (OT), where entropy-regularized OT distills the structural dark knowledge. Privileged KD with OT (PKDOT) method captures the local structures in the multimodal teacher representation by calculating a cosine similarity matrix and selects the top-k anchors to allow for sparse OT solutions, resulting in a more stable distillation process. Experiments were performed on two different problems: pain estimation on the Biovid dataset (ordinal classification) and arousal-valance prediction on the Affwild2 dataset (regression). Results show that the proposed method can outperform state-of-the-art privileged KD methods on these problems. The diversity of different modalities and fusion architectures indicates that the proposed PKDOT method is modality and model-agnostic.
The dot product self-attention (DPSA) is a fundamental component of transformers. However, scaling them to long sequences, like documents or high-resolution images, becomes prohibitively expensive due to quadratic time and memory complexities arising from the softmax operation. Kernel methods are employed to simplify computations by approximating softmax but often lead to performance drops compared to softmax attention. We propose SeTformer, a novel transformer, where DPSA is purely replaced by Self-optimal Transport (SeT) for achieving better performance and computational efficiency. SeT is based on two essential softmax properties: maintaining a non-negative attention matrix and using a nonlinear reweighting mechanism to emphasize important tokens in input sequences. By introducing a kernel cost function for optimal transport, SeTformer effectively satisfies these properties. In particular, with small and basesized models, SeTformer achieves impressive top-1 accuracies of 84.7% and 86.2% on ImageNet-1K. In object detection, SeTformer-base outperforms the FocalNet counterpart by +2.2 mAP, using 38% fewer parameters and 29% fewer FLOPs. In semantic segmentation, our base-size model surpasses NAT by +3.5 mIoU with 33% fewer parameters. SeTformer also achieves state-of-the-art results in language modeling on the GLUE benchmark. These findings highlight SeTformer's applicability in vision and language tasks.