Computer Vision and Pattern Recognition (cs.CV)

Abstract: Autonomous vehicles and robots need to operate over a wide variety of scenarios in order to complete tasks efficiently and safely. Multi-camera self-supervised monocular depth estimation from videos is a promising way to reason about the environment, as it generates metrically scaled geometric predictions from visual data without requiring additional sensors. However, most works assume well-calibrated extrinsics to fully leverage this multi-camera setup, even though accurate and efficient calibration is still a challenging problem. In this work, we introduce a novel method for extrinsic calibration that builds upon the principles of self-supervised monocular depth and ego-motion learning. Our proposed curriculum learning strategy uses monocular depth and pose estimators with velocity supervision to estimate extrinsics, and then jointly learns extrinsic calibration along with depth and pose for a set of overlapping cameras rigidly attached to a moving vehicle. Experiments on a benchmark multi-camera dataset (DDAD) demonstrate that our method enables self-calibration in various scenes robustly and efficiently compared to a traditional vision-based pose estimation pipeline. Furthermore, we demonstrate the benefits of extrinsics self-calibration as a way to improve depth prediction via joint optimization.

Abstract: Recent score-based diffusion models (SBDMs) show promising results in unpaired image-to-image translation (I2I). However, existing methods, either energy-based or statistically-based, provide no explicit form of the interfered intermediate generative distributions. This work presents a new score-decomposed diffusion model (SDDM) on manifolds to explicitly optimize the tangled distributions during image generation. SDDM derives manifolds to make the distributions of adjacent time steps separable and decompose the score function or energy guidance into an image ``denoising" part and a content ``refinement" part. To refine the image in the same noise level, we equalize the refinement parts of the score function and energy guidance, which permits multi-objective optimization on the manifold. We also leverage the block adaptive instance normalization module to construct manifolds with lower dimensions but still concentrated with the perturbed reference image. SDDM outperforms existing SBDM-based methods with much fewer diffusion steps on several I2I benchmarks.

Abstract: Due to the progression of information technology in recent years, document images have been widely disseminated in social networks. With the help of powerful image editing tools, document images are easily forged without leaving visible manipulation traces, which leads to severe issues if significant information is falsified for malicious use. Therefore, the research of document image forensics is worth further exploring. In a document image, the character with specific semantic information is most vulnerable to tampering, for which capturing the forgery traces of the character is the key to localizing the forged region in document images. Considering both character and image textures, in this paper, we propose a Character Texture Perception Network (CTP-Net) to localize the forgery of document images. Based on optical character recognition, a Character Texture Stream (CTS) is designed to capture features of text areas that are essential components of a document image. Meanwhile, texture features of the whole document image are exploited by an Image Texture Stream (ITS). Combining the features extracted from the CTS and the ITS, the CTP-Net can reveal more subtle forgery traces from document images. To overcome the challenge caused by the lack of fake document images, we design a data generation strategy that is utilized to construct a Fake Chinese Trademark dataset (FCTM). Through a series of experiments, we show that the proposed CTP-Net is able to capture tampering traces in document images, especially in text regions. Experimental results demonstrate that CTP-Net can localize multi-scale forged areas in document images and outperform the state-of-the-art forgery localization methods.

Abstract: Recently, vision Transformers (ViTs) have been actively applied to fine-grained visual recognition (FGVR). ViT can effectively model the interdependencies between patch-divided object regions through an inherent self-attention mechanism. In addition, patch selection is used with ViT to remove redundant patch information and highlight the most discriminative object patches. However, existing ViT-based FGVR models are limited to single-scale processing, and their fixed receptive fields hinder representational richness and exacerbate vulnerability to scale variability. Therefore, we propose multi-scale patch selection (MSPS) to improve the multi-scale capabilities of existing ViT-based models. Specifically, MSPS selects salient patches of different scales at different stages of a multi-scale vision Transformer (MS-ViT). In addition, we introduce class token transfer (CTT) and multi-scale cross-attention (MSCA) to model cross-scale interactions between selected multi-scale patches and fully reflect them in model decisions. Compared to previous single-scale patch selection (SSPS), our proposed MSPS encourages richer object representations based on feature hierarchy and consistently improves performance from small-sized to large-sized objects. As a result, we propose M2Former, which outperforms CNN-/ViT-based models on several widely used FGVR benchmarks.

Abstract: Referring video object segmentation (RVOS) is a task that aims to segment the target object in all video frames based on a sentence describing the object. Previous RVOS methods have achieved significant performance with densely-annotated datasets, whose construction is expensive and time-consuming. To relieve the burden of data annotation while maintaining sufficient supervision for segmentation, we propose a new annotation scheme, in which we label the frame where the object first appears with a mask and use bounding boxes for the subsequent frames. Based on this scheme, we propose a method to learn from this weak annotation. Specifically, we design a cross frame segmentation method, which uses the language-guided dynamic filters to thoroughly leverage the valuable mask annotation and bounding boxes. We further develop a bi-level contrastive learning method to encourage the model to learn discriminative representation at the pixel level. Extensive experiments and ablative analyses show that our method is able to achieve competitive performance without the demand of dense mask annotation. The code will be available at https://github.com/wangbo-zhao/WRVOS/.

Abstract: Whilst deep learning techniques have achieved excellent emotion prediction, they still require large amounts of labelled training data, which are (a) onerous and tedious to compile, and (b) prone to errors and biases. We propose Multi-Task Contrastive Learning for Affect Representation (\textbf{MT-CLAR}) for few-shot affect inference. MT-CLAR combines multi-task learning with a Siamese network trained via contrastive learning to infer from a pair of expressive facial images (a) the (dis)similarity between the facial expressions, and (b) the difference in valence and arousal levels of the two faces. We further extend the image-based MT-CLAR framework for automated video labelling where, given one or a few labelled video frames (termed \textit{support-set}), MT-CLAR labels the remainder of the video for valence and arousal. Experiments are performed on the AFEW-VA dataset with multiple support-set configurations; moreover, supervised learning on representations learnt via MT-CLAR are used for valence, arousal and categorical emotion prediction on the AffectNet and AFEW-VA datasets. The results show that valence and arousal predictions via MT-CLAR are very comparable to the state-of-the-art (SOTA), and we significantly outperform SOTA with a support-set $\approx$6\% the size of the video dataset.

Abstract: Affordance learning considers the interaction opportunities for an actor in the scene and thus has wide application in scene understanding and intelligent robotics. In this paper, we focus on contextual affordance learning, i.e., using affordance as context to generate a reasonable human pose in a scene. Existing scene-aware human pose generation methods could be divided into two categories depending on whether using pose templates. Our proposed method belongs to the template-based category, which benefits from the representative pose templates. Moreover, inspired by recent transformer-based methods, we associate each query embedding with a pose template, and use the interaction between query embeddings and scene feature map to effectively predict the scale and offsets for each pose template. In addition, we employ knowledge distillation to facilitate the offset learning given the predicted scale. Comprehensive experiments on Sitcom dataset demonstrate the effectiveness of our method.

Abstract: Anomaly detection, or outlier detection, is a crucial task in various domains to identify instances that significantly deviate from established patterns or the majority of data. In the context of autonomous driving, the identification of anomalies is particularly important to prevent safety-critical incidents, as deep learning models often exhibit overconfidence in anomalous or outlier samples. In this study, we explore different strategies for training an image semantic segmentation model with an anomaly detection module. By introducing modifications to the training stage of the state-of-the-art DenseHybrid model, we achieve significant performance improvements in anomaly detection. Moreover, we propose a simplified detector that achieves comparable results to our modified DenseHybrid approach, while also surpassing the performance of the original DenseHybrid model. These findings demonstrate the efficacy of our proposed strategies for enhancing anomaly detection in the context of autonomous driving.

Abstract: Compared to the multi-stage self-supervised multi-view stereo (MVS) method, the end-to-end (E2E) approach has received more attention due to its concise and efficient training pipeline. Recent E2E self-supervised MVS approaches have integrated third-party models (such as optical flow models, semantic segmentation models, NeRF models, etc.) to provide additional consistency constraints, which grows GPU memory consumption and complicates the model's structure and training pipeline. In this work, we propose an efficient framework for end-to-end self-supervised MVS, dubbed ES-MVSNet. To alleviate the high memory consumption of current E2E self-supervised MVS frameworks, we present a memory-efficient architecture that reduces memory usage by 43% without compromising model performance. Furthermore, with the novel design of asymmetric view selection policy and region-aware depth consistency, we achieve state-of-the-art performance among E2E self-supervised MVS methods, without relying on third-party models for additional consistency signals. Extensive experiments on DTU and Tanks&Temples benchmarks demonstrate that the proposed ES-MVSNet approach achieves state-of-the-art performance among E2E self-supervised MVS methods and competitive performance to many supervised and multi-stage self-supervised methods.

Abstract: Direct training of Spiking Neural Networks (SNNs) on neuromorphic hardware has the potential to significantly reduce the high energy consumption of Artificial Neural Networks (ANNs) training on modern computers. The biological plausibility of SNNs allows them to benefit from bio-inspired plasticity rules, such as Spike Timing-Dependent Plasticity (STDP). STDP offers gradient-free and unsupervised local learning, which can be easily implemented on neuromorphic hardware. However, relying solely on unsupervised STDP to perform classification tasks is not enough. In this paper, we propose Stabilized Supervised STDP (S2-STDP), a supervised STDP learning rule to train the classification layer of an SNN equipped with unsupervised STDP. S2-STDP integrates error-modulated weight updates that align neuron spikes with desired timestamps derived from the average firing time within the layer. Then, we introduce a training architecture called Paired Competing Neurons (PCN) to further enhance the learning capabilities of our classification layer trained with S2-STDP. PCN associates each class with paired neurons and encourages neuron specialization through intra-class competition. We evaluated our proposed methods on image recognition datasets, including MNIST, Fashion-MNIST, and CIFAR-10. Results showed that our methods outperform current supervised STDP-based state of the art, for comparable architectures and numbers of neurons. Also, the use of PCN enhances the performance of S2-STDP, regardless of the configuration, and without introducing any hyperparameters.Further analysis demonstrated that our methods exhibited improved hyperparameter robustness, which reduces the need for tuning.

Abstract: Conventional detectors suffer from performance degradation when dealing with long-tailed data due to a classification bias towards the majority head categories. In this paper, we contend that the learning bias originates from two factors: 1) the unequal competition arising from the imbalanced distribution of foreground categories, and 2) the lack of sample diversity in tail categories. To tackle these issues, we introduce a unified framework called BAlanced CLassification (BACL), which enables adaptive rectification of inequalities caused by disparities in category distribution and dynamic intensification of sample diversities in a synchronized manner. Specifically, a novel foreground classification balance loss (FCBL) is developed to ameliorate the domination of head categories and shift attention to difficult-to-differentiate categories by introducing pairwise class-aware margins and auto-adjusted weight terms, respectively. This loss prevents the over-suppression of tail categories in the context of unequal competition. Moreover, we propose a dynamic feature hallucination module (FHM), which enhances the representation of tail categories in the feature space by synthesizing hallucinated samples to introduce additional data variances. In this divide-and-conquer approach, BACL sets a new state-of-the-art on the challenging LVIS benchmark with a decoupled training pipeline, surpassing vanilla Faster R-CNN with ResNet-50-FPN by 5.8% AP and 16.1% AP for overall and tail categories. Extensive experiments demonstrate that BACL consistently achieves performance improvements across various datasets with different backbones and architectures. Code and models are available at https://github.com/Tianhao-Qi/BACL.

Abstract: Discovering ancient agricultural terraces in desert regions is important for the monitoring of long-term climate changes on the Earth's surface. However, traditional ground surveys are both costly and limited in scale. With the increasing accessibility of aerial and satellite data, machine learning techniques bear large potential for the automatic detection and recognition of archaeological landscapes. In this paper, we propose a deep semantic model fusion method for ancient agricultural terrace detection. The input data includes aerial images and LiDAR generated terrain features in the Negev desert. Two deep semantic segmentation models, namely DeepLabv3+ and UNet, with EfficientNet backbone, are trained and fused to provide segmentation maps of ancient terraces and walls. The proposed method won the first prize in the International AI Archaeology Challenge. Codes are available at https://github.com/wangyi111/international-archaeology-ai-challenge.

Abstract: Painterly image harmonization aims to insert photographic objects into paintings and obtain artistically coherent composite images. Previous methods for this task mainly rely on inference optimization or generative adversarial network, but they are either very time-consuming or struggling at fine control of the foreground objects (e.g., texture and content details). To address these issues, we propose a novel Painterly Harmonization stable Diffusion model (PHDiffusion), which includes a lightweight adaptive encoder and a Dual Encoder Fusion (DEF) module. Specifically, the adaptive encoder and the DEF module first stylize foreground features within each encoder. Then, the stylized foreground features from both encoders are combined to guide the harmonization process. During training, besides the noise loss in diffusion model, we additionally employ content loss and two style losses, i.e., AdaIN style loss and contrastive style loss, aiming to balance the trade-off between style migration and content preservation. Compared with the state-of-the-art models from related fields, our PHDiffusion can stylize the foreground more sufficiently and simultaneously retain finer content. Our code and model are available at https://github.com/bcmi/PHDiffusion-Painterly-Image-Harmonization.

Abstract: View Transformation Module (VTM), where transformations happen between multi-view image features and Bird-Eye-View (BEV) representation, is a crucial step in camera-based BEV perception systems. Currently, the two most prominent VTM paradigms are forward projection and backward projection. Forward projection, represented by Lift-Splat-Shoot, leads to sparsely projected BEV features without post-processing. Backward projection, with BEVFormer being an example, tends to generate false-positive BEV features from incorrect projections due to the lack of utilization on depth. To address the above limitations, we propose a novel forward-backward view transformation module. Our approach compensates for the deficiencies in both existing methods, allowing them to enhance each other to obtain higher quality BEV representations mutually. We instantiate the proposed module with FB-BEV, which achieves a new state-of-the-art result of 62.4\% NDS on the nuScenes test set. The code will be released at \url{https://github.com/NVlabs/FB-BEV}.

Abstract: Estimating surface normals from 3D point clouds is critical for various applications, including surface reconstruction and rendering. While existing methods for normal estimation perform well in regions where normals change slowly, they tend to fail where normals vary rapidly. To address this issue, we propose a novel approach called MSECNet, which improves estimation in normal varying regions by treating normal variation modeling as an edge detection problem. MSECNet consists of a backbone network and a multi-scale edge conditioning (MSEC) stream. The MSEC stream achieves robust edge detection through multi-scale feature fusion and adaptive edge detection. The detected edges are then combined with the output of the backbone network using the edge conditioning module to produce edge-aware representations. Extensive experiments show that MSECNet outperforms existing methods on both synthetic (PCPNet) and real-world (SceneNN) datasets while running significantly faster. We also conduct various analyses to investigate the contribution of each component in the MSEC stream. Finally, we demonstrate the effectiveness of our approach in surface reconstruction.

Abstract: Estimating 6D poses and reconstructing 3D shapes of objects in open-world scenes from RGB-depth image pairs is challenging. Many existing methods rely on learning geometric features that correspond to specific templates while disregarding shape variations and pose differences among objects in the same category. As a result, these methods underperform when handling unseen object instances in complex environments. In contrast, other approaches aim to achieve category-level estimation and reconstruction by leveraging normalized geometric structure priors, but the static prior-based reconstruction struggles with substantial intra-class variations. To solve these problems, we propose the DTF-Net, a novel framework for pose estimation and shape reconstruction based on implicit neural fields of object categories. In DTF-Net, we design a deformable template field to represent the general category-wise shape latent features and intra-category geometric deformation features. The field establishes continuous shape correspondences, deforming the category template into arbitrary observed instances to accomplish shape reconstruction. We introduce a pose regression module that shares the deformation features and template codes from the fields to estimate the accurate 6D pose of each object in the scene. We integrate a multi-modal representation extraction module to extract object features and semantic masks, enabling end-to-end inference. Moreover, during training, we implement a shape-invariant training strategy and a viewpoint sampling method to further enhance the model's capability to extract object pose features. Extensive experiments on the REAL275 and CAMERA25 datasets demonstrate the superiority of DTF-Net in both synthetic and real scenes. Furthermore, we show that DTF-Net effectively supports grasping tasks with a real robot arm.

Abstract: The confidence calibration of deep learning-based perception models plays a crucial role in their reliability. Especially in the context of autonomous driving, downstream tasks like prediction and planning depend on accurate confidence estimates. In point-wise multiclass classification tasks like sematic segmentation the model has to deal with heavy class imbalances. Due to their underrepresentation, the confidence calibration of classes with smaller instances is challenging but essential, not only for safety reasons. We propose a metric to measure the confidence calibration quality of a semantic segmentation model with respect to individual classes. It is calculated by computing sparsification curves for each class based on the uncertainty estimates. We use the classification calibration metric to evaluate uncertainty estimation methods with respect to their confidence calibration of underrepresented classes. We furthermore suggest a double use for the method to automatically find label problems to improve the quality of hand- or auto-annotated datasets.

Abstract: To evaluate perception components of an automated driving system, it is necessary to define the relevant objects. While the urban domain is popular among perception datasets, relevance is insufficiently specified for this domain. Therefore, this work adopts an existing method to define relevance in the highway domain and expands it to the urban domain. While different conceptualizations and definitions of relevance are present in literature, there is a lack of methods to validate these definitions. Therefore, this work presents a novel relevance validation method leveraging a motion prediction component. The validation leverages the idea that removing irrelevant objects should not influence a prediction component which reflects human driving behavior. The influence on the prediction is quantified by considering the statistical distribution of prediction performance across a large-scale dataset. The validation procedure is verified using criteria specifically designed to exclude relevant objects. The validation method is successfully applied to the relevance criteria from this work, thus supporting their validity.

Abstract: Depth estimation aims to predict dense depth maps. In autonomous driving scenes, sparsity of annotations makes the task challenging. Supervised models produce concave objects due to insufficient structural information. They overfit to valid pixels and fail to restore spatial structures. Self-supervised methods are proposed for the problem. Their robustness is limited by pose estimation, leading to erroneous results in natural scenes. In this paper, we propose a supervised framework termed Diffusion-Augmented Depth Prediction (DADP). We leverage the structural characteristics of diffusion model to enforce depth structures of depth models in a plug-and-play manner. An object-guided integrality loss is also proposed to further enhance regional structure integrality by fetching objective information. We evaluate DADP on three driving benchmarks and achieve significant improvements in depth structures and robustness. Our work provides a new perspective on depth estimation with sparse annotations in autonomous driving scenes.

Abstract: Recent advances in Scene Graph Generation (SGG) typically model the relationships among entities utilizing box-level features from pre-defined detectors. We argue that an overlooked problem in SGG is the coarse-grained interactions between boxes, which inadequately capture contextual semantics for relationship modeling, practically limiting the development of the field. In this paper, we take the initiative to explore and propose a generic paradigm termed Superpixel-based Interaction Learning (SIL) to remedy coarse-grained interactions at the box level. It allows us to model fine-grained interactions at the superpixel level in SGG. Specifically, (i) we treat a scene as a set of points and cluster them into superpixels representing sub-regions of the scene. (ii) We explore intra-entity and cross-entity interactions among the superpixels to enrich fine-grained interactions between entities at an earlier stage. Extensive experiments on two challenging benchmarks (Visual Genome and Open Image V6) prove that our SIL enables fine-grained interaction at the superpixel level above previous box-level methods, and significantly outperforms previous state-of-the-art methods across all metrics. More encouragingly, the proposed method can be applied to boost the performance of existing box-level approaches in a plug-and-play fashion. In particular, SIL brings an average improvement of 2.0% mR (even up to 3.4%) of baselines for the PredCls task on Visual Genome, which facilitates its integration into any existing box-level method.

Abstract: Deep Neural Network (DNN) has achieved great success on datasets of closed class set. However, new classes, like new categories of social media topics, are continuously added to the real world, making it necessary to incrementally learn. This is hard for DNN because it tends to focus on fitting to new classes while ignoring old classes, a phenomenon known as catastrophic forgetting. State-of-the-art methods rely on knowledge distillation and data replay techniques but still have limitations. In this work, we analyze the causes of catastrophic forgetting in class incremental learning, which owes to three factors: representation drift, representation confusion, and classifier distortion. Based on this view, we propose a two-stage learning framework with a fixed encoder and an incrementally updated prototype classifier. The encoder is trained with self-supervised learning to generate a feature space with high intrinsic dimensionality, thus improving its transferability and generality. The classifier incrementally learns new prototypes while retaining the prototypes of previously learned data, which is crucial in preserving the decision boundary.Our method does not rely on preserved samples of old classes, is thus a non-exemplar based CIL method. Experiments on public datasets show that our method can significantly outperform state-of-the-art exemplar-based methods when they reserved 5 examplers per class, under the incremental setting of 10 phases, by 18.24% on CIFAR-100 and 9.37% on ImageNet100.

Abstract: This is the Algonauts 2023 submission report for team "BlobGPT". Our model consists of a multi-subject linear encoding head attached to a pretrained trunk model. The multi-subject head consists of three components: (1) a shared multi-layer feature projection, (2) shared plus subject-specific low-dimension linear transformations, and (3) a shared PCA fMRI embedding. In this report, we explain these components in more detail and present some experimental results. Our code is available at https://github.com/cmi-dair/algonauts23.

Abstract: Remote sensing image change detection aims to identify the differences between images acquired at different times in the same area. It is widely used in land management, environmental monitoring, disaster assessment and other fields. Currently, most change detection methods are based on Siamese network structure or early fusion structure. Siamese structure focuses on extracting object features at different times but lacks attention to change information, which leads to false alarms and missed detections. Early fusion (EF) structure focuses on extracting features after the fusion of images of different phases but ignores the significance of object features at different times for detecting change details, making it difficult to accurately discern the edges of changed objects. To address these issues and obtain more accurate results, we propose a novel network, Triplet UNet(T-UNet), based on a three-branch encoder, which is capable to simultaneously extract the object features and the change features between the pre- and post-time-phase images through triplet encoder. To effectively interact and fuse the features extracted from the three branches of triplet encoder, we propose a multi-branch spatial-spectral cross-attention module (MBSSCA). In the decoder stage, we introduce the channel attention mechanism (CAM) and spatial attention mechanism (SAM) to fully mine and integrate detailed textures information at the shallow layer and semantic localization information at the deep layer.

Abstract: Optical Character Recognition (OCR) enables automatic text extraction from scanned or digitized text images, but it also makes it easy to pirate valuable or sensitive text from these images. Previous methods to prevent OCR piracy by distorting characters in text images are impractical in real-world scenarios, as pirates can capture arbitrary portions of the text images, rendering the defenses ineffective. In this work, we propose a novel and effective defense mechanism termed the Universal Defensive Underpainting Patch (UDUP) that modifies the underpainting of text images instead of the characters. UDUP is created through an iterative optimization process to craft a small, fixed-size defensive patch that can generate non-overlapping underpainting for text images of any size. Experimental results show that UDUP effectively defends against unauthorized OCR under the setting of any screenshot range or complex image background. It is agnostic to the content, size, colors, and languages of characters, and is robust to typical image operations such as scaling and compressing. In addition, the transferability of UDUP is demonstrated by evading several off-the-shelf OCRs. The code is available at https://github.com/QRICKDD/UDUP.

Abstract: Diffeomorphic registration has become a powerful approach for seeking a smooth and invertible spatial transformation between two coordinate systems which have been measured via the template and reference images. While the pointwise volume-preserving constraint is effective for some problems, it is too stringent for many other problems especially when the local deformations are relatively large, because it may lead to a poor large-deformation for enforcing local matching.In this paper, we propose a novel bi-variant diffeomorphic image registration model with the soft constraint of Jacobian equation, which allows local deformations to shrink and grow in a flexible range.The Jacobian determinant of the transformation is explicitly controlled by optimizing the relaxation function. To prevent deformation folding and enhance the smoothness of deformation, we not only impose a positivity constraint in optimizing the relaxation function, but also employ a regularizer to ensure the smoothness of the relaxation function.Furthermore, the positivity constraint ensures that is as close to one as possible, which helps to obtain a volume-preserving transformation on average.We further analyze the existence of the minimizer for the variational model and propose a penalty splitting method with a multilevel strategy to solve this model. Numerical experiments show that the proposed algorithm is convergent, and the positivity constraint can control the range of relative volume and not compromise registration accuracy. Moreover, the proposed model produces diffeomorphic maps for large deformation, and achieves better performance compared to the several existing registration models.

Abstract: In this study, we aim to initiate the development of Radiology Foundation Model, termed as RadFM.We consider the construction of foundational models from the perspectives of data, model design, and evaluation thoroughly. Our contribution can be concluded as follows: (i), we construct a large-scale Medical Multi-modal Dataset, MedMD, consisting of 16M 2D and 3D medical scans. To the best of our knowledge, this is the first multi-modal dataset containing 3D medical scans. (ii), We propose an architecture that enables visually conditioned generative pre-training, allowing for the integration of text input interleaved with 2D or 3D medical scans to generate response for diverse radiologic tasks. The model was initially pre-trained on MedMD and subsequently domain-specific fine-tuned on RadMD, a radiologic cleaned version of MedMD, containing 3M radiologic visual-language pairs. (iii), we propose a new evaluation benchmark that comprises five tasks, aiming to comprehensively assess the capability of foundation models in handling practical clinical problems. Our experimental results confirm that RadFM significantly outperforms existing multi-modal foundation models. The codes, data, and model checkpoint will all be made publicly available to promote further research and development in the field.

Abstract: Open-vocabulary segmentation is a challenging task requiring segmenting and recognizing objects from an open set of categories. One way to address this challenge is to leverage multi-modal models, such as CLIP, to provide image and text features in a shared embedding space, which bridges the gap between closed-vocabulary and open-vocabulary recognition. Hence, existing methods often adopt a two-stage framework to tackle the problem, where the inputs first go through a mask generator and then through the CLIP model along with the predicted masks. This process involves extracting features from images multiple times, which can be ineffective and inefficient. By contrast, we propose to build everything into a single-stage framework using a shared Frozen Convolutional CLIP backbone, which not only significantly simplifies the current two-stage pipeline, but also remarkably yields a better accuracy-cost trade-off. The proposed FC-CLIP, benefits from the following observations: the frozen CLIP backbone maintains the ability of open-vocabulary classification and can also serve as a strong mask generator, and the convolutional CLIP generalizes well to a larger input resolution than the one used during contrastive image-text pretraining. When training on COCO panoptic data only and testing in a zero-shot manner, FC-CLIP achieve 26.8 PQ, 16.8 AP, and 34.1 mIoU on ADE20K, 18.2 PQ, 27.9 mIoU on Mapillary Vistas, 44.0 PQ, 26.8 AP, 56.2 mIoU on Cityscapes, outperforming the prior art by +4.2 PQ, +2.4 AP, +4.2 mIoU on ADE20K, +4.0 PQ on Mapillary Vistas and +20.1 PQ on Cityscapes, respectively. Additionally, the training and testing time of FC-CLIP is 7.5x and 6.6x significantly faster than the same prior art, while using 5.9x fewer parameters. FC-CLIP also sets a new state-of-the-art performance across various open-vocabulary semantic segmentation datasets. Code at https://github.com/bytedance/fc-clip