WACV 2024 Papers

The Space-Time Video Super-Resolution (STVSR) task aims to enhance the visual quality of videos, by simultaneously performing video frame interpolation (VFI) and video super-resolution (VSR). However, facing the challenge of the additional temporal dimension and scale inconsistency, most existing STVSR methods are complex and inflexible in dynamically modeling different motion amplitudes. In this work, we find that choosing an appropriate processing scale achieves remarkable benefits in flow-based feature propagation. We propose a novel Scale-Adaptive Feature Aggregation (SAFA) network that adaptively selects sub-networks with different processing scales for individual samples. Experiments on four public STVSR benchmarks demonstrate that SAFA achieves state-of-the-art performance. Our SAFA network outperforms recent state-of-the-art methods such as TMNet and VideoINR by an average improvement of over 0.5dB on PSNR, while requiring less than half the number of parameters and only 1/3 computational costs. Our code will be publicly released.

Salient object detection is the task of producing a binary mask for an image that deciphers which pixels belong to the foreground object versus background. We introduce a new salient object detection dataset using images taken by people who are visually impaired who were seeking to better understand their surroundings, which we call VizWiz-SalientObject. Compared to seven existing datasets, VizWiz-SalientObject is the largest (i.e., 32,000 human-annotated images) and contains unique characteristics including a higher prevalence of text in the salient objects (i.e., in 68% of images) and salient objects that occupy a larger ratio of the images (i.e., on average, 50% coverage). We benchmarked ten modern models on our dataset. While most methods fall below human performance, struggling most for images with salient objects that are large, have less complex boundaries, and lack text as well as for lower quality images, one method one method is very close. To facilitate future extensions of this work, we publicly share the dataset at https://vizwiz.org/tasks-and-datasets/salient-object-detection.

arge-scale foundation models, such as CLIP, have demonstrated impressive zero-shot generalization performance on downstream tasks, leveraging well-designed language prompts. However, these prompt learning techniques often struggle with domain shift, limiting their generalization capabilities. In our study, we tackle this issue by proposing STYLIP, a novel approach for Domain Generalization (DG) that enhances CLIP's classification performance across domains. Our method focuses on a domain-agnostic prompt learning strategy, aiming to disentangle the visual style and content information embedded in CLIP's pre-trained vision encoder, enabling effortless adaptation to novel domains during inference. To achieve this, we introduce a set of style projectors that directly learn the domain-specific prompt tokens from the extracted multi-scale style features. These generated prompt embeddings are subsequently combined with the multi-scale visual content features learned by a content projector. The projectors are trained in a contrastive manner, utilizing CLIP's fixed vision and text backbones. Through extensive experiments conducted in five different DG settings on multiple benchmark datasets, we consistently demonstrate that STYLIP outperforms the current state-of-the-art (SOTA) methods.

We introduce the structured scene-text spotting task, which requires a scene-text OCR system to spot text in the wild according to a query regular expression. Contrary to generic scene-text OCR, structured scene-text spotting seeks to dynamically condition both detection and recognition on user-provided regular expressions. To tackle this task, we propose the Structured TExt sPotter (STEP), a model that exploits the provided text structure to guide the OCR process. STEP is able to deal with regular expressions that contain spaces and it is not bound to detection at word-level granularity. Our approach enables accurate zero-shot structured text spotting in a wide variety of real-world reading scenarios and is solely trained on publicly available data. To demonstrate the effectiveness of our approach, we introduce a new challenging test dataset that contains several types of out-of-vocabulary structured text, reflecting important reading applications such as weight information, serial numbers, license plates etc. We demonstrate that STEP can provide specialized OCR performance on demand in all tested scenarios. The code and test dataset are released at https://github.com/CVC-DAG/STEP.

Despite significant progress in semi-supervised learning for image object detection, several key issues are yet to be addressed for video object detection: (1) Achieving good performance for supervised video object detection greatly depends on the availability of annotated frames. (2) Despite having large inter-frame correlations in a video, collecting annotations for a large number of frames per video is expensive, time-consuming, and often redundant. (3) Existing semi-supervised techniques on static images can hardly exploit the temporal motion dynamics inherently present in videos. In this paper, we introduce SSVOD, an end-to-end semi-supervised video object detection framework that exploits motion dynamics of videos to utilize large-scale unlabeled frames with sparse annotations. To selectively assemble robust pseudo-labels across groups of frames, we introduce flow-warped predictions from nearby frames for temporal-consistency estimation. In particular, we introduce cross-IoU and cross-divergence based selection methods over a set of estimated predictions to include robust pseudo-labels for bounding boxes and class labels, respectively. To strike a balance between confirmation bias and uncertainty noise in pseudo-labels, we propose confidence threshold based combination of hard and soft pseudo-labels. Our method achieves significant performance improvements over existing methods on ImageNet-VID, Epic-KITCHENS, and YouTube-VIS datasets. Codes are available at https://github.com/enyac-group/SSVOD.git.

Reconstructing 3D geometry from unoriented point clouds can benefit many downstream tasks. % Recent shape modeling methods mostly adopt implicit neural representation to fit a signed distance field (SDF) and optimize the network by unsigned supervision. % However, these methods occasionally have difficulty in finding the coarse shape for complicated objects, especially suffering from the "ghost" surfaces (i.e., fake surfaces that should not exist). % To guide the network quickly fit the coarse shape, we propose to utilize the signed supervision in regions that are obviously outside the object and can be easily determined, resulting in our semi-signed supervision. % To better recover high-fidelity details, a novel loss-based region sampling strategy and a progressive positional encoding (PE) method are applied to prioritize the optimization towards underfitting and complicated regions. % Specifically, we voxelize and partition the object space into sign-known and sign-uncertain regions, in which different supervisions are applied. % Besides, we adaptively adjust the sampling rate of each voxel according to the tracked reconstruction loss, so that the network can focus more on the complicated under-fitting regions. % We conduct extensive experiments to demonstrate that our method achieves state-of-the-art performance compared to the existing fitting-based methods and comparable performance to learning-based methods on multiple datasets. % The code is publicly available at https://github.com/Runsong123/SSP.

We consider the problem of cross-sensor domain adaptation in the context of LiDAR-based 3D object detection and propose Stationary Object Aggregation Pseudo-labelling (SOAP) to generate high quality pseudo-labels for stationary objects. In contrast to the current state-of-the-art in-domain practice of aggregating just a few input scans, SOAP aggregates entire sequences of point clouds at the input level to reduce the sensor domain gap. Then, by means of what we call quasi-stationary training and spatial consistency post-processing, the SOAP model generates accurate pseudo-labels for stationary objects, closing a minimum of 30.3% domain gap compared to few-frame detectors. Our results also show that state-of-the-art domain adaptation approaches can achieve even greater performance in combination with SOAP, in both the unsupervised and semi-supervised settings.

Learning from set-structured data is an essential problem with many applications in machine learning and computer vision. This paper focuses on non-parametric and data-independent learning from set-structured data using approximate nearest neighbor (ANN) solutions, particularly locality-sensitive hashing. We consider the problem of set retrieval from an input set query. Such a retrieval problem requires: 1) an efficient mechanism to calculate the distances/dissimilarities between sets, and 2) an appropriate data structure for fast nearest-neighbor search. To that end, we propose to use Sliced-Wasserstein embedding as a computationally efficient set-2-vector operator that enables downstream ANN, with theoretical guarantees. The set elements are treated as samples from an unknown underlying distribution, and the Sliced-Wasserstein distance is used to compare sets. We demonstrate the effectiveness of our algorithm, denoted as Set Locality Sensitive Hashing (SLoSH), on various set retrieval datasets and compare our proposed embedding with standard set embedding approaches, including Generalized Mean (GeM) embedding/pooling, Featurewise Sort Pooling (FSPool), Covariance Pooling, and Wasserstein embedding and show consistent improvement in retrieval results.

The availability of well-curated datasets has driven the success of Machine Learning (ML) models. Despite greater access to earth observation data in agriculture, there is a scarcity of curated and labelled datasets, which limits the potential of its use in training ML models for remote sensing (RS) in agriculture. To this end, we introduce a first-of-its-kind dataset called SICKLE, which constitutes a time-series of multi-resolution imagery from 3 distinct satellites: Landsat-8, Sentinel-1 and Sentinel-2. Our dataset constitutes multi-spectral, thermal and microwave sensors during January 2018 - March 2021 period. We construct each temporal sequence by considering the cropping practices followed by farmers primarily engaged in paddy cultivation in the Cauvery Delta region of Tamil Nadu, India; and annotate the corresponding imagery with key cropping parameters at multiple resolutions (i.e. 3m, 10m and 30m). Our dataset comprises 2, 370 season-wise samples from 388 unique plots, having an average size of 0.38 acres, for classifying 21 crop types across 4 districts in the Delta, which amounts to approximately 209, 000 satellite images. Out of the 2, 370 samples, 351 paddy samples from 145 plots are annotated with multiple crop parameters; such as the variety of paddy, its growing season and productivity in terms of per-acre yields. Ours is also one among the first studies that consider the growing season activities pertinent to crop phenology (spans sowing, transplanting and harvesting dates) as parameters of interest. We benchmark SICKLE on three tasks: crop type, crop phenology (sowing, transplanting, harvesting), and yield prediction.

In the field of 3D scene understanding, 3D scene graphs have emerged as a new scene representation that combines geometric and semantic information about objects and their relationships. However, learning semantic 3D scene graphs in a fully supervised manner is inherently difficult as it requires not only object-level annotations but also relationship labels. While pre-training approaches have helped to boost the performance of many methods in various fields, pre-training for 3D scene graph prediction has received little attention. Furthermore, we find in this paper that classical contrastive point cloud-based pre-training approaches are ineffective for 3D scene graph learning. To this end, we present SGRec3D, a novel self-supervised pre-training method for 3D scene graph prediction. We propose to reconstruct the 3D input scene from a graph bottleneck as a pretext task. Pre-training SGRec3D does not require object relationship labels, making it possible to exploit large-scale 3D scene understanding datasets, which were off-limits for 3D scene graph learning before. Our experiments demonstrate that in contrast to recent point cloud-based pre-training approaches, our proposed pre-training improves the 3D scene graph prediction considerably, which results in SOTA performance, outperforming other 3D scene graph models by +10% on object prediction and +4% on relationship prediction. Additionally, we show that only using a small subset of 10% labeled data during fine-tuning is sufficient to outperform the same model without pre-training.

Transformer architecture is a de-facto standard for modeling global dependency in long sequences. However, quadratic space and time complexity for self-attention prohibits transformers from scaling to extremely long sequences (> 10k). Low-rank decomposition as a non-negative matrix factorization (NMF) of self-attention demonstrates remarkable performance in linear space and time complexity with strong theoretical guarantees. However, our analysis reveals that NMF-based works struggle to capture the rich spatio-temporal visual cues scattered across the long sequences resulting from egocentric lifelogs. To capture such cues, we propose a novel attention mechanism named SEMantic Attention (SEMA), which factorizes the self-attention matrix into a semantically meaningful subspace. We demonstrate SEMA in a representation learning setting, aiming to recover activity patterns in extremely long (weeks-long) egocentric lifelogs using a novel self-supervised training pipeline. Compared to the current state-of-the-art, we report significant improvement in terms of (NMI, AMI, and F-Score) for EgoRoutine, UTE, and Epic Kitchens datasets. Furthermore, to underscore the efficacy of SEMA, we extend its application to conventional video tasks such as online action detection, video recognition, and action localization.

Most existing portrait matting models either require expensive auxiliary information or try to decompose the task into sub-tasks that are usually resource-hungry. These challenges limit its application on low-power computing devices. In addition, mobile networks tend to be less powerful than those cumbersome ones in feature representation mining. In this paper, we propose an extremely efficient portrait matting model via self-distillation (SDNet), that aims to provide a solution to performing accurate and effective portrait matting with limited computing resources. Our SDNet contains only 2M parameters, 2.2% of the parameters of MGM, and 1.5% of that of Matteformer. We introduce the training pipeline of self-distillation that can improve our lightweight baseline model without any parameter addition, network modification, or over-parameterized teacher models which need well-pretraining. Extensive experiments demonstrate the effectiveness of our self-distillation method and the lightweight SDNet network. Our SDNet outperforms the state-of-the-art (SOTA) lightweight approaches on both synthetic and real-world images.

We propose Subject-Conditional Relation Detection SCoRD, where conditioned on an input subject, the goal is to predict all its relations to other objects in a scene along with their locations. Based on the Open Images dataset, we propose a challenging OIv6-SCoRD benchmark such that the training and testing splits have a distribution shift in terms of the occurrence statistics of <subject, relation, object> triplets. To solve this problem, we propose an auto-regressive model that given a subject, it predicts its relations, objects, and object locations by casting this output as a sequence of tokens. First, we show that previous scene-graph prediction methods fail to produce as exhaustive an enumeration of relation-object pairs when conditioned on a subject on this benchmark. Particularly, we obtain a recall@3 of 83.8% for our relation-object predictions compared to the 49.75% obtained by a recent scene graph detector. Then, we show improved generalization on both relation-object and object-box predictions by leveraging during training relation-object pairs obtained automatically from textual captions and for which no object-box annotations are available. Particularly, for <subject, relation, object> triplets for which no object locations are available during training, we are able to obtain a recall@3 of 33.80% for relation-object pairs and 26.75% for their box locations.

Pulmonary embolism (PE) is a prevalent lung disease that can lead to right ventricular hypertrophy and failure in severe cases, ranking second in severity only to myocardial infarction and sudden death. Pulmonary artery CT angiography (CTPA) is a widely used diagnostic method for PE. However, PE detection presents challenges in clinical practice due to limitations in imaging technology. CTPA can produce noises similar to PE, making confirmation of its presence time-consuming and prone to overdiagnosis. Nevertheless, the traditional segmentation method of PE can not fully consider the hierarchical structure of features, local and global spatial features of PE CT images. In this paper, we propose an automatic PE segmentation method called SCUNet++ (Swin Conv UNet++). This method incorporates multiple fusion dense skip connections between the encoder and decoder, utilizing the Swin Transformer as the encoder. And fuses features of different scales in the decoder subnetwork to compensate for spatial information loss caused by the inevitable downsampling in Swin-UNet or other state-of-the-art methods, effectively solving the above problem. We provide a theoretical analysis of this method in detail and validate it on publicly available PE CT image datasets FUMPE and CAD-PE. The experimental results indicate that our proposed method achieved a Dice similarity coefficient (DSC) of 83.47% and a Hausdorff distance 95th percentile (HD95) of 3.83 on the FUMPE dataset, as well as a DSC of 83.42% and an HD95 of 5.10 on the CAD-PE dataset. These findings demonstrate that our method exhibits strong performance in PE segmentation tasks, potentially enhancing the accuracy of automatic segmentation of PE and providing a powerful diagnostic tool for clinical physicians. Our source code and new FUMPE dataset are available at https://github.com/JustlfC03/SCUNet-plusplus.

Multiple Instance learning (MIL) models have been extensively used in pathology to predict biomarkers and risk-stratify patients from gigapixel-sized images. Machine learning problems in medical imaging often deal with rare diseases, making it important for these models to work in a label-imbalanced setting. In pathology images, there is another level of imbalance, where given a positively labeled Whole Slide Image (WSI), only a fraction of pixels within it contribute to the positive label. This compounds the severity of imbalance and makes imbalanced classification in pathology challenging. Furthermore, these imbalances can occur in out-of-distribution (OOD) datasets when the models are deployed in the real-world. We leverage the idea that decoupling feature and classifier learning can lead to improved decision boundaries for label imbalanced datasets. To this end, we investigate the integration of supervised contrastive learning with multiple instance learning (SC-MIL). Specifically, we propose a joint-training MIL framework in the presence of label imbalance that progressively transitions from learning bag-level representations to optimal classifier learning. We perform experiments with different imbalance settings for two well-studied problems in cancer pathology: subtyping of non-small cell lung cancer and subtyping of renal cell carcinoma. SC-MIL provides large and consistent improvements over other techniques on both in-distribution (ID) and OOD held-out sets across multiple imbalanced settings.

Computer vision has become increasingly prevalent in solving real-world problems across diverse domains, including smart agriculture, fishery, and livestock management. These applications may not require processing many image frames per second, leading practitioners to use single board computers (SBCs). Although many lightweight networks have been developed for "mobile/edge" devices, they primarily target smartphones with more powerful processors and not SBCs with the low-end CPUs. This paper introduces a CNN-ViT hybrid network called SBCFormer, which achieves high accuracy and fast computation on such low-end CPUs. The hardware constraints of these CPUs make the Transformer's attention mechanism preferable to convolution. However, using attention on low-end CPUs presents a challenge: high-resolution internal feature maps demand excessive computational resources, but reducing their resolution results in the loss of local image details. SBCFormer introduces an architectural design to address this issue. As a result, SBCFormer achieves the highest trade-off between accuracy and speed on a Raspberry Pi 4 Model B with an ARM-Cortex A72 CPU. For the first time, it achieves an ImageNet-1K top-1 accuracy of around 80% at a speed of 1.0 frame/sec on the SBC. Code is available at https://github.com/xyongLu/SBCFormer.

We propose a straightforward yet highly effective few-shot fine-tuning strategy for adapting the Segment Anything (SAM) to anatomical segmentation tasks in medical images. Our novel approach revolves around reformulating the mask decoder within SAM, leveraging few-shot embeddings derived from a limited set of labeled images (few-shot collection) as prompts for querying anatomical objects captured in image embeddings. This innovative reformulation greatly reduces the need for time-consuming online user interactions for labeling volumetric images, such as exhaustively marking points and bounding boxes to provide prompts slice by slice. With our method, users can manually segment a few 2D slices offline, and the embeddings of these annotated image regions serve as effective prompts for online segmentation tasks. Our method prioritizes the efficiency of the fine-tuning process by exclusively training the mask decoder through caching mechanisms while keeping the image encoder frozen. Importantly, this approach is not limited to volumetric medical images, but can generically be applied to any 2D/3D segmentation task. To thoroughly evaluate our method, we conducted extensive validation on four datasets, covering six anatomical segmentation tasks across two modalities. Furthermore, we conducted a comparative analysis of different prompting options within SAM and the fully-supervised nnU-Net. The results demonstrate the superior performance of our method compared to SAM employing only point prompts (50% improvement in IoU) and performs on-par with fully supervised methods whilst reducing the requirement of labeled data by at least an order of magnitude.

Crop detection is integral for precision agriculture applications such as automated yield estimation or fruit picking. However, crop detection, e.g., apple detection in orchard environments remains challenging due to a lack of large-scale datasets and the small relative size of the crops in the image. In this work, we address these challenges by reformulating the apple detection task in a semi-supervised manner. To this end, we provide the large, high-resolution dataset MAD comprising 105 labeled images with 14,667 annotated apple instances and 4,440 unlabeled images. Utilizing this dataset, we also propose a novel Semi-Supervised Small Apple Detection system S3AD based on contextual attention and selective tiling to improve the challenging detection of small apples, while limiting the computational overhead. We conduct an extensive evaluation on MAD and the MSU dataset, showing that S3AD substantially outperforms strong fully-supervised baselines, including several small object detection systems, by up to 14.9%. Additionally, we exploit the detailed annotations of our dataset w.r.t. apple properties to analyze the influence of relative size or level of occlusion on the results of various systems, quantifying current challenges.

Appearance-based gaze estimation has been actively studied in recent years. However, its generalization performance for unseen head poses is still a significant limitation for existing methods. This work proposes a generalizable multi-view gaze estimation task and a cross-view feature fusion method to address this issue. In addition to paired images, our method takes the relative rotation matrix between two cameras as additional input. The proposed network learns to extract rotatable feature representation by using relative rotation as a constraint and adaptively fuses the rotatable features via stacked fusion modules. This simple yet efficient approach significantly improves generalization performance under unseen head poses without significantly increasing computational cost. The model can be trained with random combinations of cameras without fixing the positioning and can generalize to unseen camera pairs during inference. Through experiments using multiple datasets, we demonstrate the advantage of the proposed method over baseline methods, including state-of-the-art domain generalization approaches.