WACV 2024 Papers

We propose few-shot generative models of skeleton-based human actions on limited samples of the target domain. We exploit large public datasets as a source of motion variations by introducing novel cross-domain and entropy regularization losses that effectively transfer the diversity of the motions contained in the source to the target domain. First, target samples are divided into patches, which are a set of short motion clips. For each patch, we search for a reference motion from the source dataset that is similar to the patch. Next, in adversarial training, our cross-domain regularization encourages the generated sequences to resemble the reference motion at the patch level. Entropy regularization prevents mode collapse by forcing the generator to follow the distribution of the source dataset. Experiments are performed on public datasets where we utilize three action classes from NTU RGB+D 120 as the target and all data of 60 action classes in NTU RGB+D as the source. Ten samples for each target action class, 30 in total, are selected as target data. The results demonstrate that data augmented with the proposed method improve recognition accuracy by 28 % using a ST-GCN classifier.

Event classification in images plays a vital role in multimedia analysis especially with the prevalence of fake news on social media and the Web. The majority of approaches for event classification rely on large sets of labeled training data. However, image labels for fine-grained event instances (e.g., 2016 Summer Olympics) can be sparse, incorrect, ambiguous, etc. A few approaches have addressed the lack of labeled data for event classification but cover only few events. Moreover, vision-language models that allow for zero-shot and few-shot classification with prompting have not yet been extensively exploited. In this paper, we propose four different techniques to create hard prompts including knowledge graph information from Wikidata and Wikipedia as well as an ensemble approach for zero-shot event classification. We also integrate prompt learning for state-of-the-art vision-language models to address few-shot event classification. Experimental results on six benchmarks including a new dataset comprising event instances from various domains, such as politics and natural disasters, show that our proposed approaches require much fewer training images than supervised baselines and the state-of-the-art while achieving better results.

We study a new highly-practical problem setting that enables resource-constrained edge devices to adapt a pre-trained model to their local data distributions. Recognizing that device's data are likely to come from multiple latent domains that include a mixture of unlabelled domain-relevant and domain-irrelevant examples, we focus on the comparatively under-studied problem of latent domain adaptation. Considering limitations of edge devices, we aim to only use a pre-trained model and adapt it in a feed-forward way, without using back-propagation and without access to the source data. Modelling these realistic constraints bring us to the novel and practically important problem setting of feed-forward latent domain adaptation. Our solution is to meta-learn a network capable of embedding the mixed-relevance target dataset and dynamically adapting inference for target examples using cross-attention. The resulting framework leads to consistent improvements over strong ERM baselines. We also show that our framework sometimes even improves on the upper bound of domain-supervised adaptation, where only domain-relevant instances are provided for adaptation. This suggests that human annotated domain labels may not always be optimal, and raises the possibility of doing better through automated instance selection.

We investigate the reasons for lower performance of transformer based pedestrian detection models compared to convolutional neural network (CNN) based ones. CNN models generate dense pedestrian proposals, refine each proposal individually, and follow it up with non-maximal-suppression (NMS) to generate sparse predictions. In contrast, transformer models select one proposal per ground-truth (GT) pedestrian box and backpropagate positive gradient from them. All other proposals, many of them highly similar to the selected ones, are passed negative gradient. Though this leads to sparse predictions, obviating the need of NMS, the arbitrary selection of one among many similar proposals, hinders effective training, and lower accuracy of pedestrian detection. To mitigate the problem, instead of commonly used Kuhn-Munkres matching algorithm, we propose Min-cost-flow based formulation, and incorporate constraints such as, each ground truth box is matched to atleast one proposal, and many equally good proposals can be matched to a single ground truth box. We propose first transformer based pedestrian detection model incorporating our matching algorithm. Extensive experiments reveal that our approach achieves a miss rate (lower is better) of 3.7 / 17.4 / 21.8 / 8.3 / 2.0 on Eurocity / TJU-traffic / TJU-campus / Cityperson / Caltech datasets compared to 4.7 / 18.7 / 24.8 / 8.5 / 3.1 by the current SOTA. Code is available at https://ajayshastry08.github.io/flow_matcher

Super-resolution (SR) techniques have recently been proposed to upscale the outputs of neural radiance fields (NeRF) and generate high-quality images with enhanced inference speeds. However, existing NeRF+SR methods increase training overhead by using extra input features, loss functions, or expensive training procedures such as knowledge distillation. In this paper, we aim to leverage SR for efficiency gains without costly training or architectural changes. Specifically, we build a simple NeRF+SR pipeline that directly combines existing modules, and we propose a lightweight augmentation technique, random patch sampling, for training. Compared to existing NeRF+SR methods, our pipeline mitigates the SR computing overhead and can be trained up to 23x faster, making it feasible to run on consumer devices such as the Apple MacBook. Experiments show that our pipeline can upscale NeRF outputs by 2-4x while maintaining high quality, increasing inference speeds by up to 18x on an NVIDIA V100 GPU and 12.8x on an M1 Pro chip. We conclude that SR can be a simple but effective technique for improving the efficiency of NeRF models for consumer devices.

In recent years, language-driven artistic style transfer has emerged as a new type of style transfer technique, eliminating the need for a reference style image by using natural language descriptions of the style. The first model to achieve this, called CLIPstyler, has demonstrated impressive stylisation results. However, its lengthy optimisation procedure at runtime for each query limits its suitability for many practical applications. In this work, we present FastCLIPstyler, a generalised text-based image style transfer model capable of stylising images in a single forward pass for arbitrary text inputs. Furthermore, we introduce EdgeCLIPstyler, a lightweight model designed for compatibility with resource-constrained devices. Through quantitative and qualitative comparisons with state-of-the-art approaches, we demonstrate that our models achieve superior stylisation quality based on measurable metrics while offering significantly improved runtime efficiency, particularly on edge devices.

Most face identification approaches employ a Siamese neural network to compare two images at the image embedding level. Yet, this technique can be subject to occlusion (e.g., faces with masks or sunglasses) and out-of-distribution data. DeepFace-EMD (Phan et al. 2022) reaches state-of-the-art accuracy on out-of-distribution data by first comparing two images at the image level, and then at the patch level. Yet, its later patch-wise re-ranking stage admits a large O(n^3 log n) time complexity (for n patches in an image) due to the optimal transport optimization. In this paper, we propose a novel, 2-image Vision Transformers (ViTs) that compares two images at the patch level using cross-attention. After training on 2M pairs of images on CASIA Webface (Yi et al. 2014), our model performs at a comparable accuracy as DeepFace-EMD on out-of-distribution data, yet at an inference speed more than twice as fast as DeepFace-EMD (Phan et al. 2022). In addition, via a human study, our model shows promising explainability through the visualization of cross-attention. We believe our work can inspire more explorations in using ViTs for face identification.

In this work, we introduce our method of outdoor scene relighting for Neural Radiance Fields (NeRF) named Sun-aligned Relighting TensoRF (SR-TensoRF). SR-TensoRF offers a lightweight and rapid pipeline aligned with the sun, thereby achieving a simplified workflow that eliminates the need for environment maps. Our sun-alignment strategy is motivated by the insight that shadows, unlike viewpoint-dependent albedo, are determined by light direction. We directly use the sun direction as an input during shadow generation, simplifying the requirements of the inference process significantly. Moreover, SR-TensoRF leverages the training efficiency of TensoRF by incorporating our proposed cubemap concept, resulting in notable acceleration in both training and rendering processes compared to existing methods.

Using diffusion models to solve inverse problems is a growing field of research. Current methods assume the degradation to be known and provide impressive results in terms of restoration quality and diversity. In this work, we leverage the efficiency of those models to jointly estimate the restored image and unknown parameters of the degradation model such as blur kernel. In particular, we designed an algorithm based on the well-known Expectation-Minimization (EM) estimation method and diffusion models. Our method alternates between approximating the expected log-likelihood of the inverse problem using samples drawn from a diffusion model and a maximization step to estimate unknown model parameters. For the maximization step, we also introduce a novel blur kernel regularization based on a Plug & Play denoiser. Diffusion models are long to run, thus we provide a fast version of our algorithm. Extensive experiments on blind image deblurring demonstrate the effectiveness of our method when compared to other state-of-the-art approaches.

Whole-body biometric recognition is an important area of research due to its vast applications in law enforcement, border security, and surveillance. This paper presents the end-to-end design, development and evaluation of FarSight, an innovative software system designed for whole-body (fusion of face, gait and body shape) biometric recognition. FarSight accepts videos from elevated platforms and drones as input and outputs a candidate list of identities from a gallery. The system is designed to address several challenges, including (i) low-quality imagery, (ii) large yaw and pitch angles, (iii) robust feature extraction to accommodate large intra-person variabilities and large inter-person similarities, and (iv) the large domain gap between training and test sets. FarSight combines the physics of imaging and deep learning models to enhance image restoration and biometric feature encoding. We test FarSight's effectiveness using the newly acquired IARPA Biometric Recognition and Identification at Altitude and Range (BRIAR) dataset. Notably, FarSight demonstrated a substantial performance increase on the BRIAR dataset, with gains of +11.82% Rank-20 identification and +11.3% TAR@1%FAR.

Recent face presentation attack detection (PAD) leverages domain adaptation (DA) and domain generalization (DG) techniques to address performance degradation on unknown domains. However, DA-based PAD methods require access to unlabeled target data, while most DG-based PAD solutions rely on a priori, i.e., known domain labels. Moreover, most DA-/DG-based methods are computationally intensive, demanding complex model architectures and/or multi-stage training processes. This paper proposes to model face PAD as a compound DG task from a causal perspective, linking it to model optimization. We excavate the causal factors hidden in the high-level representation via counterfactual intervention. Moreover, we introduce a class-guided MixStyle to enrich feature-level data distribution within classes instead of focusing on domain information. Both class-guided MixStyle and counterfactual intervention components introduce no extra trainable parameters and negligible computational resources. Extensive cross-dataset and analytic experiments demonstrate the effectiveness and efficiency of our method compared to state-of-the-art PADs. The implementation and the trained weights are publicly available.

Conditional GANs are frequently used for manipulating the attributes of face images, such as expression, hairstyle, pose, or age. Even though the state-of-the-art models successfully modify the requested attributes, they simultaneously modify other important characteristics of the image, such as a person's identity. In this paper, we focus on solving this problem by introducing PluGeN4Faces, a plugin to StyleGAN, which explicitly disentangles face attributes from a person's identity. Our key idea is to perform training on images retrieved from movie frames, where a given person appears in various poses and with different attributes. By applying a type of contrastive loss, we encourage the model to group images of the same person in similar regions of latent space. Our experiments demonstrate that the modifications of face attributes performed by PluGeN4Faces are significantly less invasive on the remaining characteristics of the image than in the existing state-of-the-art models.

We introduce FacadeNet, a deep learning approach for synthesizing building facade images from diverse viewpoints. Our method employs a conditional GAN, taking a single view of a facade along with the desired viewpoint information and generates an image of the facade from the distinct viewpoint. To precisely modify view-dependent elements like windows and doors while preserving the structure of view-independent components such as walls, we introduce a selective editing module. This module leverages image embeddings extracted from a pretrained vision transformer Our experiments demonstrated state-of-the-art performance on building facade generation, surpassing alternative methods.

This paper proposes a generic multi-object tracking (MOT) algorithm that is robust to unexpected motion changes for generic objects. Deep learning has dramatically been improving MOT performances. Nevertheless, state-of-the-art tracking algorithms are still sensitive to unexpected motion changes and the generic object target beyond person tracking. This is because standard MOT benchmark datasets such as MOT17 mainly consist of persons in a crowd, often lacking unexpected shape and motion changes; thus, these issues have yet to be focused on. We propose a simple-yet-effective MOT framework that can dynamically improve tracking continuity by associating each target based on adaptively modified motion states. The keys are 1) to represent the target motions using multiple motion states that have weak correlations with each other and 2) to modify those states that have the lowest similarity to past states as outliers. Our approach can overwhelmingly improve trajectory continuity and robustness to unexpected motion changes for generic objects. Comprehensive experiments have confirmed that our framework is comparable to existing state-of-the-art methods on a standard dataset and outperforms those algorithms on the GMOT dataset with an overall 2% improvement in IDF1, a measure of tracking continuity.

Training fingerprint recognition models using synthetic data has recently gained increased attention in the biometric community as it alleviates the dependency on sensitive personal data. Existing approaches for fingerprint generation are limited in their ability to generate diverse impressions of the same finger, a key property for providing effective data for training recognition models. To address this gap, we present FPGAN-Control, an identity preserving image generation framework which enables control over the fingerprint's image appearance (e.g., fingerprint type, acquisition device, pressure level) of generated fingerprints. We introduce a novel appearance loss that encourages disentanglement between the fingerprint's identity and appearance properties. In our experiments, we used the publicly available NIST SD302 (N2N) dataset for training the FPGAN-Control model. We demonstrate the merits of FPGAN-Control, both quantitatively and qualitatively, in terms of identity preservation level, degree of appearance control, and low synthetic-to-real domain gap. Finally, training recognition models using only synthetic datasets generated by FPGAN-Control lead to recognition accuracies that are on par or even surpass models trained using real data. To the best of our knowledge, this is the first work to demonstrate this.

Surface reconstruction from multi-view images is a challenging task, with solutions often requiring a large number of sampled images with high overlap. We seek to develop a method for few-view reconstruction, for the case of the human foot. To solve this task, we must extract rich geometric cues from RGB images, before carefully fusing them into a final 3D object. Our FOUND approach tackles this, with 4 main contributions: (i) SynFoot, a synthetic dataset of 50,000 photorealistic foot images, paired with ground truth surface normals and keypoints; (ii) an uncertainty-aware surface normal predictor trained on our synthetic dataset; (iii) an optimization scheme for fitting a generative foot model to a series of images; and (iv) a benchmark dataset of calibrated images and high resolution ground truth geometry. We show that our normal predictor outperforms all off-the-shelf equivalents significantly on real images, and our optimization scheme outperforms state-of-the-art photogrammetry pipelines, especially for a few-view setting. We release our synthetic dataset and baseline 3D scans to the research community.

Unsupervised Open-Vocabulary Semantic Segmentation aims to segment an image into regions referring to an arbitrary set of concepts described by text, without relying on dense annotations that are available only for a subset of the categories. Previous works relied on inducing pixel-level alignment in a multi-modal space through contrastive training over vast corpora of image-caption pairs. However, representing a semantic category solely through its textual embedding is insufficient to encompass the wide-ranging variability in the visual appearances of the images associated with that category. In this paper, we propose FOSSIL, a pipeline that enables a self-supervised backbone to perform open-vocabulary segmentation relying only on the visual modality. In particular, we decouple the task into two components: (1) we leverage text-conditioned diffusion models to generate a large collection of visual embeddings, starting from a set of captions. These can be retrieved at inference time to obtain a support set of references for the set of textual concepts. Further, (2) we exploit self-supervised dense features to partition the image into semantically coherent regions. We demonstrate that our approach provides strong performance on different semantic segmentation datasets, without requiring any additional training.

Vision Transformers (ViT) have recently demonstrated success across a myriad of computer vision tasks. However, their elevated computational demands pose significant challenges for real-world deployment. While low-rank approximation stands out as a renowned method to reduce computational loads, efficiently automating the target rank selection in ViT remains a challenge. Drawing from the notable similarity and alignment between the processes of rank selection and One-Shot NAS, we introduce FLORA, an end-to-end automatic framework based on NAS. To overcome the design challenge of supernet posed by vast search space, FLORA employs a low-rank aware candidate filtering strategy. This method adeptly identifies and eliminates underperforming candidates, effectively alleviating potential undertraining and interference among subnetworks. To further enhance the quality of low-rank supernets, we design a low-rank specific training paradigm. First, we propose weight inheritance to construct supernet and enable gradient sharing among low-rank modules. Secondly, we adopt low-rank aware sampling to strategically allocate training resources, taking into account inherited information from pre-trained models. Empirical results underscore FLORA's efficacy. With our method, a more fine-grained rank configuration can be generated automatically and yield up to 33% extra FLOPs reduction compared to a simple uniform configuration. More specific, FLORA-DeiT-B/FLORA-Swin-B can save up to 55%/42% FLOPs almost without performance degradtion. Importantly, FLORA boasts both versatility and orthogonality, offering an extra 21%-26% FLOPs reduction when integrated with leading compression techniques or compact hybrid structures. Our code is publicly available at https://github.com/shadowpa0327/FLORA.

Neural 3D implicit representations learn priors that are useful for diverse applications, such as single- or multiple-view 3D reconstruction. A major downside of existing approaches while rendering an image is that they require evaluating the network multiple times per camera ray so that the high computational time forms a bottleneck for downstream applications. We address this problem by introducing a novel neural scene representation that we call the directional distance function (DDF). To this end, we learn a signed distance function (SDF) along with our DDF model to represent a class of shapes. Specifically, our DDF is defined on the unit sphere and predicts the distance to the surface along any given direction. Therefore, our DDF allows rendering images with just a single network evaluation per camera ray. Based on our DDF, we present a novel fast algorithm (FIRe) to reconstruct 3D shapes given a posed depth map. We evaluate our proposed method on 3D reconstruction from single-view depth images, where we empirically show that our algorithm reconstructs 3D shapes more accurately and it is more than 15 times faster (per iteration) than competing methods.