The architecture of transformers, which recently witness booming applications in vision tasks, has pivoted against the widespread convolutional paradigm. Relying on the tokenization process that splits inputs into multiple tokens, transformers are capable of extracting their pairwise relationships using self-attention. While being the stemming building block of transformers, what makes for a good tokenizer has not been well understood in computer vision. In this work, we investigate this uncharted problem from an information trade-off perspective. In addition to unifying and understanding existing structural modifications, our derivation leads to better design strategies for vision tokenizers. The proposed Modulation across Tokens (MoTo) incorporates inter-token modeling capability through normalization. Furthermore, a regularization objective TokenProp is embraced in the standard training regime. Through extensive experiments on various transformer architectures, we observe both improved performance and intriguing properties of these two plug-and-play designs with negligible computational overhead. These observations further indicate the importance of the commonly-omitted designs of tokenizers in vision transformer.

Pre-trained large language models can efficiently interpolate human-written prompts in a natural way. Multitask prompted learning can help generalization through a diverse set of tasks at once, thus enhancing the potential for more effective downstream fine-tuning. To perform efficient multitask-inference in the same batch, parameter-efficient fine-tuning methods such as prompt tuning have been proposed. However, the existing prompt tuning methods may lack generalization. We propose SPT, a semi-parametric prompt tuning method for multitask prompted learning. The novel component of SPT is a memory bank from where memory prompts are retrieved based on discrete prompts. Extensive experiments, such as (i) fine-tuning a full language model with SPT on 31 different tasks from 8 different domains and evaluating zero-shot generalization on 9 heldout datasets under 5 NLP task categories and (ii) pretraining SPT on the GLUE datasets and evaluating fine-tuning on the SuperGLUE datasets, demonstrate effectiveness of SPT.

To build Video Question Answering (VideoQA) systems capable of assisting humans in daily activities, seeking answers from long-form videos with diverse and complex events is a must. Existing multi-modal VQA models achieve promising performance on images or short video clips, especially with the recent success of large-scale multi-modal pre-training. However, when extending these methods to long-form videos, new challenges arise. On the one hand, using a dense video sampling strategy is computationally prohibitive. On the other hand, methods relying on sparse sampling struggle in scenarios where multi-event and multi-granularity visual reasoning are required. In this work, we introduce a new model named Multi-modal Iterative Spatial-temporal Transformer (MIST) to better adapt pre-trained models for long-form VideoQA. Specifically, MIST decomposes traditional dense spatial-temporal self-attention into cascaded segment and region selection modules that adaptively select frames and image regions that are closely relevant to the question itself. Visual concepts at different granularities are then processed efficiently through an attention module. In addition, MIST iteratively conducts selection and attention over multiple layers to support reasoning over multiple events. The experimental results on four VideoQA datasets, including AGQA, NExT-QA, STAR, and Env-QA, show that MIST achieves state-of-the-art performance and is superior at computation efficiency and interpretability.

Multimodal image-text models have shown remarkable performance in the past few years. However, evaluating their robustness against distribution shifts is crucial before adopting them in real-world applications. In this paper, we investigate the robustness of 9 popular open-sourced image-text models under common perturbations on five tasks (image-text retrieval, visual reasoning, visual entailment, image captioning, and text-to-image generation). In particular, we propose several new multimodal robustness benchmarks by applying 17 image perturbation and 16 text perturbation techniques on top of existing datasets. We observe that multimodal models are not robust to image and text perturbations, especially to image perturbations. Among the tested perturbation methods, character-level perturbations constitute the most severe distribution shift for text, and zoom blur is the most severe shift for image data. We also introduce two new robustness metrics (MMI and MOR) for proper evaluations of multimodal models. We hope our extensive study sheds light on new directions for the development of robust multimodal models.

Referring image segmentation aims at localizing all pixels of the visual objects described by a natural language sentence. Previous works learn to straightforwardly align the sentence embedding and pixel-level embedding for highlighting the referred objects, but ignore the semantic consistency of pixels within the same object, leading to incomplete masks and localization errors in predictions. To tackle this problem, we propose CoupAlign, a simple yet effective multi-level visual-semantic alignment method, to couple sentence-mask alignment with word-pixel alignment to enforce object mask constraint for achieving more accurate localization and segmentation. Specifically, the Word-Pixel Alignment (WPA) module performs early fusion of linguistic and pixel-level features in intermediate layers of the vision and language encoders. Based on the word-pixel aligned embedding, a set of mask proposals are generated to hypothesize possible objects. Then in the Sentence-Mask Alignment (SMA) module, the masks are weighted by the sentence embedding to localize the referred object, and finally projected back to aggregate the pixels for the target. To further enhance the learning of the two alignment modules, an auxiliary loss is designed to contrast the foreground and background pixels. By hierarchically aligning pixels and masks with linguistic features, our CoupAlign captures the pixel coherence at both visual and semantic levels, thus generating more accurate predictions. Extensive experiments on popular datasets (e.g., RefCOCO and G-Ref) show that our method achieves consistent improvements over state-of-the-art methods, e.g., about 2% oIoU increase on the validation and testing set of RefCOCO. Especially, CoupAlign has remarkable ability in distinguishing the target from multiple objects of the same class.

Monocular depth estimation is a challenging problem on which deep neural networks have demonstrated great potential. However, depth maps predicted by existing deep models usually lack fine-grained details due to the convolution operations and the down-samplings in networks. We find that increasing input resolution is helpful to preserve more local details while the estimation at low resolution is more accurate globally. Therefore, we propose a novel depth map fusion module to combine the advantages of estimations with multi-resolution inputs. Instead of merging the low- and high-resolution estimations equally, we adopt the core idea of Poisson fusion, trying to implant the gradient domain of high-resolution depth into the low-resolution depth. While classic Poisson fusion requires a fusion mask as supervision, we propose a self-supervised framework based on guided image filtering. We demonstrate that this gradient-based composition performs much better at noisy immunity, compared with the state-of-the-art depth map fusion method. Our lightweight depth fusion is one-shot and runs in real-time, making our method 80X faster than a state-of-the-art depth fusion method. Quantitative evaluations demonstrate that the proposed method can be integrated into many fully convolutional monocular depth estimation backbones with a significant performance boost, leading to state-of-the-art results of detail enhancement on depth maps.

Utilities of knowledge graphs (KGs) depend on their qualities. A KG that is of poor quality not only has little applicability but also leads to some unexpected errors. Therefore, quality evaluation for KGs is crucial and indispensable. Existing methods design many quality dimensions and calculate metrics in the corresponding dimensions based on details (i.e., raw data and graph structures) of KGs for evaluation. However, there are two major issues. On one hand, they consider the details as public information, which exposes the raw data and graph structures. These details are strictly confidential because they involve commercial privacy or others in practice. On the other hand, the existing methods focus on how much knowledge KGs have rather than KGs' practicability. To address the above problems, we propose a knowledge graph quality evaluation framework under incomplete information (QEII). The quality evaluation problem is transformed into an adversarial game, and the relative quality is evaluated according to the winner and loser. Participants of the game are KGs, and the adversarial gameplay is to question and answer (Q&A). In the QEII, we generate and train a question model and an answer model for each KG. The question model of a KG first asks a certain number of questions to the other KG. Then it evaluates the answers returned by the answer model of the other KG and outputs a percentage score. The relative quality is evaluated by the scores, which measures the ability to apply knowledge. Q&A messages are the only information that KGs exchange, without exposing any raw data and graph structure. Experimental results on two pairs of KGs demonstrate that, comparing with baselines, the QEII realizes a reasonable quality evaluation from the perspective of third-party evaluators under incomplete information.

Cross-view geo-localization aims to spot images of the same location shot from two platforms, e.g., the drone platform and the satellite platform. Existing methods usually focus on optimizing the distance between one embedding with others in the feature space, while neglecting the redundancy of the embedding itself. In this paper, we argue that the low redundancy is also of importance, which motivates the model to mine more diverse patterns. To verify this point, we introduce a simple yet effective regularization, i.e., Dynamic Weighted Decorrelation Regularization (DWDR), to explicitly encourage networks to learn independent embedding channels. As the name implies, DWDR regresses the embedding correlation coefficient matrix to a sparse matrix, i.e., the identity matrix, with dynamic weights. The dynamic weights are applied to focus on still correlated channels during training. Besides, we propose a cross-view symmetric sampling strategy, which keeps the example balance between different platforms. Albeit simple, the proposed method has achieved competitive results on three large-scale benchmarks, i.e., University-1652, CVUSA and CVACT. Moreover, under the harsh circumstance, e.g., the extremely short feature of 64 dimensions, the proposed method surpasses the baseline model by a clear margin.

Earth observation, aiming at monitoring the state of planet Earth using remote sensing data, is critical for improving our daily lives and living environment. With a growing number of satellites in orbit, an increasing number of datasets with diverse sensors and research domains are being published to facilitate the research of the remote sensing community. In this paper, we present a comprehensive review of more than 400 publicly published datasets, including applications like land use/cover, change/disaster monitoring, scene understanding, agriculture, climate change, and weather forecasting. We systematically analyze these Earth observation datasets with respect to five aspects volume, bibliometric analysis, resolution distributions, research domains, and the correlation between datasets. Based on the dataset attributes, we propose to measure, rank, and select datasets to build a new benchmark for model evaluation. Furthermore, a new platform for Earth observation, termed EarthNets, is released as a means of achieving a fair and consistent evaluation of deep learning methods on remote sensing data. EarthNets supports standard dataset libraries and cutting-edge deep learning models to bridge the gap between the remote sensing and machine learning communities. Based on this platform, extensive deep learning methods are evaluated on the new benchmark. The insightful results are beneficial to future research. The platform and dataset collections are publicly available at https://earthnets.github.io/.

Models should be able to adapt to unseen data during test-time to avoid performance drops caused by inevitable distribution shifts in real-world deployment scenarios. In this work, we tackle the practical yet challenging test-time adaptation (TTA) problem, where a model adapts to the target domain without accessing the source data. We propose a simple recipe called \textit{Data-efficient Prompt Tuning} (DePT) with two key ingredients. First, DePT plugs visual prompts into the vision Transformer and only tunes these source-initialized prompts during adaptation. We find such parameter-efficient finetuning can efficiently adapt the model representation to the target domain without overfitting to the noise in the learning objective. Second, DePT bootstraps the source representation to the target domain by memory bank-based online pseudo-labeling. A hierarchical self-supervised regularization specially designed for prompts is jointly optimized to alleviate error accumulation during self-training. With much fewer tunable parameters, DePT demonstrates not only state-of-the-art performance on major adaptation benchmarks VisDA-C, ImageNet-C, and DomainNet-126, but also superior data efficiency, i.e., adaptation with only 1\% or 10\% data without much performance degradation compared to 100\% data. In addition, DePT is also versatile to be extended to online or multi-source TTA settings.