The concepts of overfitting and generalization are vital for evaluating machine learning models. In this work, we show that the popular Recall@K metric depends on the number of classes in the dataset, which limits its ability to estimate generalization. To fix this issue, we propose a new metric, which measures retrieval performance, and, unlike Recall@K, estimates generalization. We apply the proposed metric to popular image retrieval methods and provide new insights about deep metric learning generalization.
Deep Metric Learning (DML) models rely on strong representations and similarity-based measures with specific loss functions. Proxy-based losses have shown great performance compared to pair-based losses in terms of convergence speed. However, proxies that are assigned to different classes may end up being closely located in the embedding space and hence having a hard time to distinguish between positive and negative items. Alternatively, they may become highly correlated and hence provide redundant information with the model. To address these issues, we propose a novel approach that introduces Soft Orthogonality (SO) constraint on proxies. The constraint ensures the proxies to be as orthogonal as possible and hence control their positions in the embedding space. Our approach leverages Data-Efficient Image Transformer (DeiT) as an encoder to extract contextual features from images along with a DML objective. The objective is made of the Proxy Anchor loss along with the SO regularization. We evaluate our method on four public benchmarks for category-level image retrieval and demonstrate its effectiveness with comprehensive experimental results and ablation studies. Our evaluations demonstrate the superiority of our proposed approach over state-of-the-art methods by a significant margin.
Deep metric learning (DML) based methods have been found very effective for content-based image retrieval (CBIR) in remote sensing (RS). For accurately learning the model parameters of deep neural networks, most of the DML methods require a high number of annotated training images, which can be costly to gather. To address this problem, in this paper we present an annotation cost efficient active learning (AL) method (denoted as ANNEAL). The proposed method aims to iteratively enrich the training set by annotating the most informative image pairs as similar or dissimilar, while accurately modelling a deep metric space. This is achieved by two consecutive steps. In the first step the pairwise image similarity is modelled based on the available training set. Then, in the second step the most uncertain and diverse (i.e., informative) image pairs are selected to be annotated. Unlike the existing AL methods for CBIR, at each AL iteration of ANNEAL a human expert is asked to annotate the most informative image pairs as similar/dissimilar. This significantly reduces the annotation cost compared to annotating images with land-use/land cover class labels. Experimental results show the effectiveness of our method. The code of ANNEAL is publicly available at https://git.tu-berlin.de/rsim/ANNEAL.
Pre-trained Vision-Language Models (VLMs) utilizing extensive image-text paired data have demonstrated unprecedented image-text association capabilities, achieving remarkable results across various downstream tasks. A critical challenge is how to make use of existing large-scale pre-trained VLMs, which are trained on common objects, to perform the domain-specific transfer for accomplishing domain-related downstream tasks. A critical challenge is how to make use of existing large-scale pre-trained VLMs, which are trained on common objects, to perform the domain-specific transfer for accomplishing domain-related downstream tasks. In this paper, we propose a new framework that includes the Domain pre-trained Vision-Language Model (DVLM), bridging the gap between the General Vision-Language Model (GVLM) and domain-specific downstream tasks. Moreover, we present an image-text paired dataset in the field of remote sensing (RS), RS5M, which has 5 million RS images with English descriptions. The dataset is obtained from filtering publicly available image-text paired datasets and captioning label-only RS datasets with pre-trained VLM. These constitute the first large-scale RS image-text paired dataset. Additionally, we fine-tuned the CLIP model and tried several Parameter-Efficient Fine-Tuning methods on RS5M to implement the DVLM. Experimental results show that our proposed dataset is highly effective for various tasks, and our model GeoRSCLIP improves upon the baseline or previous state-of-the-art model by $3\%\sim20\%$ in Zero-shot Classification (ZSC), $3\%\sim6\%$ in Remote Sensing Cross-Modal Text-Image Retrieval (RSCTIR) and $4\%\sim5\%$ in Semantic Localization (SeLo) tasks. Dataset and models have been released in: \url{https://github.com/om-ai-lab/RS5M}.
The robustness of multimodal deep learning models to realistic changes in the input text is critical for their applicability to important tasks such as text-to-image retrieval and cross-modal entailment. To measure robustness, several existing approaches edit the text data, but do so without leveraging the cross-modal information present in multimodal data. Information from the visual modality, such as color, size, and shape, provide additional attributes that users can include in their inputs. Thus, we propose cross-modal attribute insertions as a realistic perturbation strategy for vision-and-language data that inserts visual attributes of the objects in the image into the corresponding text (e.g., "girl on a chair" to "little girl on a wooden chair"). Our proposed approach for cross-modal attribute insertions is modular, controllable, and task-agnostic. We find that augmenting input text using cross-modal insertions causes state-of-the-art approaches for text-to-image retrieval and cross-modal entailment to perform poorly, resulting in relative drops of 15% in MRR and 20% in $F_1$ score, respectively. Crowd-sourced annotations demonstrate that cross-modal insertions lead to higher quality augmentations for multimodal data than augmentations using text-only data, and are equivalent in quality to original examples. We release the code to encourage robustness evaluations of deep vision-and-language models: https://github.com/claws-lab/multimodal-robustness-xmai.
The ability to predict the performance of a query before results are returned has been a longstanding challenge in Information Retrieval (IR) systems. Inspired by this task, we introduce, in this paper, a novel task called "Prompt Performance Prediction" (PPP) that aims to predict the performance of a prompt, before obtaining the actual generated images. We demonstrate the plausibility of our task by measuring the correlation coefficient between predicted and actual performance scores across: three datasets containing pairs of prompts and generated images AND three art domain datasets of real images and real user appreciation ratings. Our results show promising performance prediction capabilities, suggesting potential applications for optimizing user prompts.
Visual retrieval tasks such as image retrieval and person re-identification (Re-ID) aim at effectively and thoroughly searching images with similar content or the same identity. After obtaining retrieved examples, re-ranking is a widely adopted post-processing step to reorder and improve the initial retrieval results by making use of the contextual information from semantically neighboring samples. Prevailing re-ranking approaches update distance metrics and mostly rely on inefficient crosscheck set comparison operations while computing expanded neighbors based distances. In this work, we present an efficient re-ranking method which refines initial retrieval results by updating features. Specifically, we reformulate re-ranking based on Graph Convolution Networks (GCN) and propose a novel Graph Convolution based Re-ranking (GCR) for visual retrieval tasks via feature propagation. To accelerate computation for large-scale retrieval, a decentralized and synchronous feature propagation algorithm which supports parallel or distributed computing is introduced. In particular, the plain GCR is extended for cross-camera retrieval and an improved feature propagation formulation is presented to leverage affinity relationships across different cameras. It is also extended for video-based retrieval, and Graph Convolution based Re-ranking for Video (GCRV) is proposed by mathematically deriving a novel profile vector generation method for the tracklet. Without bells and whistles, the proposed approaches achieve state-of-the-art performances on seven benchmark datasets from three different tasks, i.e., image retrieval, person Re-ID and video-based person Re-ID.
Large-scale visual localization systems continue to rely on 3D point clouds built from image collections using structure-from-motion. While the 3D points in these models are represented using local image features, directly matching a query image's local features against the point cloud is challenging due to the scale of the nearest-neighbor search problem. Many recent approaches to visual localization have thus proposed a hybrid method, where first a global (per image) embedding is used to retrieve a small subset of database images, and local features of the query are matched only against those. It seems to have become common belief that global embeddings are critical for said image-retrieval in visual localization, despite the significant downside of having to compute two feature types for each query image. In this paper, we take a step back from this assumption and propose Constrained Approximate Nearest Neighbors (CANN), a joint solution of k-nearest-neighbors across both the geometry and appearance space using only local features. We first derive the theoretical foundation for k-nearest-neighbor retrieval across multiple metrics and then showcase how CANN improves visual localization. Our experiments on public localization benchmarks demonstrate that our method significantly outperforms both state-of-the-art global feature-based retrieval and approaches using local feature aggregation schemes. Moreover, it is an order of magnitude faster in both index and query time than feature aggregation schemes for these datasets. Code: \url{https://github.com/google-research/google-research/tree/master/cann}
We present MOFI, Manifold OF Images, a new vision foundation model designed to learn image representations from noisy entity annotated images. MOFI differs from previous work in two key aspects: (i) pre-training data, and (ii) training recipe. Regarding data, we introduce a new approach to automatically assign entity labels to images from noisy image-text pairs. Our approach involves employing a named entity recognition model to extract entities from the alt-text, and then using a CLIP model to select the correct entities as labels of the paired image. It's a simple, cost-effective method that can scale to handle billions of web-mined image-text pairs. Through this method, we have created Image-to-Entities (I2E), a new dataset with 1 billion images and 2 million distinct entities, covering rich visual concepts in the wild. Building upon the I2E dataset, we study different training recipes like supervised pre-training, contrastive pre-training, and multi-task learning. For contrastive pre-training, we treat entity names as free-form text, and further enrich them with entity descriptions. Experiments show that supervised pre-training with large-scale fine-grained entity labels is highly effective for image retrieval tasks, and multi-task training further improves the performance. The final MOFI model achieves 86.66% mAP on the challenging GPR1200 dataset, surpassing the previous state-of-the-art performance of 72.19% from OpenAI's CLIP model. Further experiments on zero-shot and linear probe image classification also show that MOFI outperforms a CLIP model trained on the original image-text data, demonstrating the effectiveness of the I2E dataset in learning strong image representations. We release our code and model weights at https://github.com/apple/ml-mofi.
We argue that there are many notions of 'similarity' and that models, like humans, should be able to adapt to these dynamically. This contrasts with most representation learning methods, supervised or self-supervised, which learn a fixed embedding function and hence implicitly assume a single notion of similarity. For instance, models trained on ImageNet are biased towards object categories, while a user might prefer the model to focus on colors, textures or specific elements in the scene. In this paper, we propose the GeneCIS ('genesis') benchmark, which measures models' ability to adapt to a range of similarity conditions. Extending prior work, our benchmark is designed for zero-shot evaluation only, and hence considers an open-set of similarity conditions. We find that baselines from powerful CLIP models struggle on GeneCIS and that performance on the benchmark is only weakly correlated with ImageNet accuracy, suggesting that simply scaling existing methods is not fruitful. We further propose a simple, scalable solution based on automatically mining information from existing image-caption datasets. We find our method offers a substantial boost over the baselines on GeneCIS, and further improves zero-shot performance on related image retrieval benchmarks. In fact, though evaluated zero-shot, our model surpasses state-of-the-art supervised models on MIT-States. Project page at https://sgvaze.github.io/genecis/.
We present an approach for multimodal pathology image search, using dynamic time warping (DTW) on Variational Autoencoder (VAE) latent space that is fed into a ranked choice voting scheme to retrieve multiplexed immunofluorescent imaging (mIF) that is most similar to a query H&E slide. Through training the VAE and applying DTW, we align and compare mIF and H&E slides. Our method improves differential diagnosis and therapeutic decisions by integrating morphological H&E data with immunophenotyping from mIF, providing clinicians a rich perspective of disease states. This facilitates an understanding of the spatial relationships in tissue samples and could revolutionize the diagnostic process, enhancing precision and enabling personalized therapy selection. Our technique demonstrates feasibility using colorectal cancer and healthy tonsil samples. An exhaustive ablation study was conducted on a search engine designed to explore the correlation between multiplexed Immunofluorescence (mIF) and Hematoxylin and Eosin (H&E) staining, in order to validate its ability to map these distinct modalities into a unified vector space. Despite extreme class imbalance, the system demonstrated robustness and utility by returning similar results across various data features, which suggests potential for future use in multimodal histopathology data analysis.
The emergence of cross-modal foundation models has introduced numerous approaches grounded in text-image retrieval. However, on some domain-specific retrieval tasks, these models fail to focus on the key attributes required. To address this issue, we propose a self-enhancement framework, A^{3}R, based on the CLIP-ViT/G-14, one of the largest cross-modal models. First, we perform an Attribute Augmentation strategy to enrich the textual description for fine-grained representation before model learning. Then, we propose an Adaption Re-ranking method to unify the representation space of textual query and candidate images and re-rank candidate images relying on the adapted query after model learning. The proposed framework is validated to achieve a salient improvement over the baseline and other teams' solutions in the cross-modal image retrieval track of the 1st foundation model challenge without introducing any additional samples. The code is available at \url{https://github.com/CapricornGuang/A3R}.
In this paper, we consider the problem of composed image retrieval (CIR), it aims to train a model that can fuse multi-modal information, e.g., text and images, to accurately retrieve images that match the query, extending the user's expression ability. We make the following contributions: (i) we initiate a scalable pipeline to automatically construct datasets for training CIR model, by simply exploiting a large-scale dataset of image-text pairs, e.g., a subset of LAION-5B; (ii) we introduce a transformer-based adaptive aggregation model, TransAgg, which employs a simple yet efficient fusion mechanism, to adaptively combine information from diverse modalities; (iii) we conduct extensive ablation studies to investigate the usefulness of our proposed data construction procedure, and the effectiveness of core components in TransAgg; (iv) when evaluating on the publicly available benckmarks under the zero-shot scenario, i.e., training on the automatically constructed datasets, then directly conduct inference on target downstream datasets, e.g., CIRR and FashionIQ, our proposed approach either performs on par with or significantly outperforms the existing state-of-the-art (SOTA) models. Project page: https://code-kunkun.github.io/ZS-CIR/
Stickers have become a ubiquitous part of modern-day communication, conveying complex emotions through visual imagery. To facilitate the development of more powerful algorithms for analyzing stickers, we propose a large-scale Chinese sticker dataset, namely Sticker820K, which consists of 820k image-text pairs. Each sticker has rich and high-quality textual annotations, including descriptions, optical characters, emotional labels, and style classifications. Although vision-language tasks in the domain of natural images have been well studied, directly applying the those models, such as CLIP, to sticker data is not an optimal solution due to the discrepant nature between natural and emotive image data. Therefore, we propose StickerCLIP as a benchmark model on the Sticker820K dataset. For the text-to-image retrieval task, our StickerCLIP demonstrates strong superiority over the CLIP, which achieves an absolute gain of 66.0\% in mean recall on the Sticker820K test set. Additionally, we endeavor to extend the recently popularized LLM by means of prompt tuning, integrating its ability for sticker retrieval and allowing users to retrieve stickers through instructions. We validate the feasibility of this method, demonstrating the immense potential of prompt tuning in expanding LLM abilities while not affecting the quality of upstream tasks.
Composed image retrieval extends content-based image retrieval systems by enabling users to search using reference images and captions that describe their intention. Despite great progress in developing image-text compositors to extract discriminative visual-linguistic features, we identify a hitherto overlooked issue, triplet ambiguity, which impedes robust feature extraction. Triplet ambiguity refers to a type of semantic ambiguity that arises between the reference image, the relative caption, and the target image. It is mainly due to the limited representation of the annotated text, resulting in many noisy triplets where multiple visually dissimilar candidate images can be matched to an identical reference pair (i.e., a reference image + a relative caption). To address this challenge, we propose the Consensus Network (Css-Net), inspired by the psychological concept that groups outperform individuals. Css-Net comprises two core components: (1) a consensus module with four diverse compositors, each generating distinct image-text embeddings, fostering complementary feature extraction and mitigating dependence on any single, potentially biased compositor; (2) a Kullback-Leibler divergence loss that encourages learning of inter-compositor interactions to promote consensual outputs. During evaluation, the decisions of the four compositors are combined through a weighting scheme, enhancing overall agreement. On benchmark datasets, particularly FashionIQ, Css-Net demonstrates marked improvements. Notably, it achieves significant recall gains, with a 2.77% increase in R@10 and 6.67% boost in R@50, underscoring its competitiveness in addressing the fundamental limitations of existing methods.
The performance of neural networks in content-based image retrieval (CBIR) is highly influenced by the chosen loss (objective) function. The majority of objective functions for neural models can be divided into metric learning and statistical learning. Metric learning approaches require a pair mining strategy that often lacks efficiency, while statistical learning approaches are not generating highly compact features due to their indirect feature optimization. To this end, we propose a novel repeller-attractor loss that falls in the metric learning paradigm, yet directly optimizes for the L2 metric without the need of generating pairs. Our loss is formed of three components. One leading objective ensures that the learned features are attracted to each designated learnable class anchor. The second loss component regulates the anchors and forces them to be separable by a margin, while the third objective ensures that the anchors do not collapse to zero. Furthermore, we develop a more efficient two-stage retrieval system by harnessing the learned class anchors during the first stage of the retrieval process, eliminating the need of comparing the query with every image in the database. We establish a set of four datasets (CIFAR-100, Food-101, SVHN, and Tiny ImageNet) and evaluate the proposed objective in the context of few-shot and full-set training on the CBIR task, by using both convolutional and transformer architectures. Compared to existing objective functions, our empirical evidence shows that the proposed objective is generating superior and more consistent results.
Sketch-Based Image Retrieval (SBIR) is a crucial task in multimedia retrieval, where the goal is to retrieve a set of images that match a given sketch query. Researchers have already proposed several well-performing solutions for this task, but most focus on enhancing embedding through different approaches such as triplet loss, quadruplet loss, adding data augmentation, and using edge extraction. In this work, we tackle the problem from various angles. We start by examining the training data quality and show some of its limitations. Then, we introduce a Relative Triplet Loss (RTL), an adapted triplet loss to overcome those limitations through loss weighting based on anchors similarity. Through a series of experiments, we demonstrate that replacing a triplet loss with RTL outperforms previous state-of-the-art without the need for any data augmentation. In addition, we demonstrate why batch normalization is more suited for SBIR embeddings than l2-normalization and show that it improves significantly the performance of our models. We further investigate the capacity of models required for the photo and sketch domains and demonstrate that the photo encoder requires a higher capacity than the sketch encoder, which validates the hypothesis formulated in [34]. Then, we propose a straightforward approach to train small models, such as ShuffleNetv2 [22] efficiently with a marginal loss of accuracy through knowledge distillation. The same approach used with larger models enabled us to outperform previous state-of-the-art results and achieve a recall of 62.38% at k = 1 on The Sketchy Database [30].
Textual scene graph parsing has become increasingly important in various vision-language applications, including image caption evaluation and image retrieval. However, existing scene graph parsers that convert image captions into scene graphs often suffer from two types of errors. First, the generated scene graphs fail to capture the true semantics of the captions or the corresponding images, resulting in a lack of faithfulness. Second, the generated scene graphs have high inconsistency, with the same semantics represented by different annotations. To address these challenges, we propose a novel dataset, which involves re-annotating the captions in Visual Genome (VG) using a new intermediate representation called FACTUAL-MR. FACTUAL-MR can be directly converted into faithful and consistent scene graph annotations. Our experimental results clearly demonstrate that the parser trained on our dataset outperforms existing approaches in terms of faithfulness and consistency. This improvement leads to a significant performance boost in both image caption evaluation and zero-shot image retrieval tasks. Furthermore, we introduce a novel metric for measuring scene graph similarity, which, when combined with the improved scene graph parser, achieves state-of-the-art (SOTA) results on multiple benchmark datasets for the aforementioned tasks. The code and dataset are available at https://github.com/zhuang-li/FACTUAL .
In computer vision, finding correct point correspondence among images plays an important role in many applications, such as image stitching, image retrieval, visual localization, etc. Most of the research works focus on the matching of local feature before a sampling method is employed, such as RANSAC, to verify initial matching results via repeated fitting of certain global transformation among the images. However, incorrect matches may still exist. Thus, a novel sampling scheme, Pentagon-Match (PMatch), is proposed in this work to verify the correctness of initially matched keypoints using pentagons randomly sampled from them. By ensuring shape and location of these pentagons are view-invariant with various evaluations of cross-ratio (CR), incorrect matches of keypoint can be identified easily with homography estimated from correctly matched pentagons. Experimental results show that highly accurate estimation of homography can be obtained efficiently for planar scenes of the HPatches dataset, based on keypoint matching results provided by LoFTR. Besides, accurate outlier identification for the above matching results and possible extension of the approach for multi-plane situation are also demonstrated.
Recent vision-language models have achieved tremendous advances. However, their computational costs are also escalating dramatically, making model acceleration exceedingly critical. To pursue more efficient vision-language Transformers, this paper introduces Cross-Guided Ensemble of Tokens (CrossGET), a general acceleration framework for vision-language Transformers. This framework adaptively combines tokens in real-time during inference, significantly reducing computational costs while maintaining high performance. CrossGET features two primary innovations: 1) Cross-Guided Matching and Ensemble. CrossGET leverages cross-modal guided token matching and ensemble to effectively utilize cross-modal information, achieving wider applicability across both modality-independent models, e.g., CLIP, and modality-dependent ones, e.g., BLIP2. 2) Complete-Graph Soft Matching. CrossGET introduces an algorithm for the token-matching mechanism, ensuring reliable matching results while facilitating parallelizability and high efficiency. Extensive experiments have been conducted on various vision-language tasks, such as image-text retrieval, visual reasoning, image captioning, and visual question answering. The performance on both classic multimodal architectures and emerging multimodal LLMs demonstrates the framework's effectiveness and versatility. The code is available at https://github.com/sdc17/CrossGET.