Visual Representation Learning with Transformer: A Sequence-to-Sequence Perspective

Visual representation learning is the key of solving various vision problems. Relying on the seminal grid structure priors, convolutional neural networks (CNNs) have been the de facto standard architectures of most deep vision models. For instance, classical semantic segmentation methods often adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated (i.e., atrous) convolutions or inserting attention modules. However, the FCN-based architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating visual representation learning generally as a sequence-to-sequence prediction task. Specifically, we deploy a pure Transformer to encode an image as a sequence of patches, without local convolution and resolution reduction. With the global context modeled in every layer of the Transformer, stronger visual representation can be learned for better tackling vision tasks. In particular, our segmentation model, termed as SEgmentation TRansformer (SETR), excels on ADE20K (50.28% mIoU, the first position in the test leaderboard on the day of submission), Pascal Context (55.83% mIoU) and reaches competitive results on Cityscapes. Further, we formulate a family of Hierarchical Local-Global (HLG) Transformers characterized by local attention within windows and global-attention across windows in a hierarchical and pyramidal architecture. Extensive experiments show that our method achieves appealing performance on a variety of visual recognition tasks (e.g., image classification, object detection and instance segmentation and semantic segmentation).

HunYuan_tvr for Text-Video Retrivial

Text-Video Retrieval plays an important role in multi-modal understanding and has attracted increasing attention in recent years. Most existing methods focus on constructing contrastive pairs between whole videos and complete caption sentences, while ignoring fine-grained cross-modal relationships, e.g., short clips and phrases or single frame and word. In this paper, we propose a novel method, named HunYuan\_tvr, to explore hierarchical cross-modal interactions by simultaneously exploring video-sentence, clip-phrase, and frame-word relationships. Considering intrinsic semantic relations between frames, HunYuan\_tvr first performs self-attention to explore frame-wise correlations and adaptively clusters correlated frames into clip-level representations. Then, the clip-wise correlation is explored to aggregate clip representations into a compact one to describe the video globally. In this way, we can construct hierarchical video representations for frame-clip-video granularities, and also explore word-wise correlations to form word-phrase-sentence embeddings for the text modality. Finally, hierarchical contrastive learning is designed to explore cross-modal relationships,~\emph{i.e.,} frame-word, clip-phrase, and video-sentence, which enables HunYuan\_tvr to achieve a comprehensive multi-modal understanding. Further boosted by adaptive label denosing and marginal sample enhancement, HunYuan\_tvr obtains new state-of-the-art results on various benchmarks, e.g., Rank@1 of 55.0%, 57.8%, 29.7%, 52.1%, and 57.3% on MSR-VTT, MSVD, LSMDC, DiDemo, and ActivityNet respectively.

Clustering by the Probability Distributions from Extreme Value Theory

Clustering is an essential task to unsupervised learning. It tries to automatically separate instances into coherent subsets. As one of the most well-known clustering algorithms, k-means assigns sample points at the boundary to a unique cluster, while it does not utilize the information of sample distribution or density. Comparably, it would potentially be more beneficial to consider the probability of each sample in a possible cluster. To this end, this paper generalizes k-means to model the distribution of clusters. Our novel clustering algorithm thus models the distributions of distances to centroids over a threshold by Generalized Pareto Distribution (GPD) in Extreme Value Theory (EVT). Notably, we propose the concept of centroid margin distance, use GPD to establish a probability model for each cluster, and perform a clustering algorithm based on the covering probability function derived from GPD. Such a GPD k-means thus enables the clustering algorithm from the probabilistic perspective. Correspondingly, we also introduce a naive baseline, dubbed as Generalized Extreme Value (GEV) k-means. GEV fits the distribution of the block maxima. In contrast, the GPD fits the distribution of distance to the centroid exceeding a sufficiently large threshold, leading to a more stable performance of GPD k-means. Notably, GEV k-means can also estimate cluster structure and thus perform reasonably well over classical k-means. Thus, extensive experiments on synthetic datasets and real datasets demonstrate that GPD k-means outperforms competitors. The github codes are released in https://github.com/sixiaozheng/EVT-K-means.

NMS-Loss: Learning with Non-Maximum Suppression for Crowded Pedestrian Detection

Non-Maximum Suppression (NMS) is essential for object detection and affects the evaluation results by incorporating False Positives (FP) and False Negatives (FN), especially in crowd occlusion scenes. In this paper, we raise the problem of weak connection between the training targets and the evaluation metrics caused by NMS and propose a novel NMS-Loss making the NMS procedure can be trained end-to-end without any additional network parameters. Our NMS-Loss punishes two cases when FP is not suppressed and FN is wrongly eliminated by NMS. Specifically, we propose a pull loss to pull predictions with the same target close to each other, and a push loss to push predictions with different targets away from each other. Experimental results show that with the help of NMS-Loss, our detector, namely NMS-Ped, achieves impressive results with Miss Rate of 5.92% on Caltech dataset and 10.08% on CityPersons dataset, which are both better than state-of-the-art competitors.

Incrementally Zero-Shot Detection by an Extreme Value Analyzer

Human beings not only have the ability to recognize novel unseen classes, but also can incrementally incorporate the new classes to existing knowledge preserved. However, zero-shot learning models assume that all seen classes should be known beforehand, while incremental learning models cannot recognize unseen classes. This paper introduces a novel and challenging task of Incrementally Zero-Shot Detection (IZSD), a practical strategy for both zero-shot learning and class-incremental learning in real-world object detection. An innovative end-to-end model -- IZSD-EVer was proposed to tackle this task that requires incrementally detecting new classes and detecting the classes that have never been seen. Specifically, we propose a novel extreme value analyzer to detect objects from old seen, new seen, and unseen classes, simultaneously. Additionally and technically, we propose two innovative losses, i.e., background-foreground mean squared error loss alleviating the extreme imbalance of the background and foreground of images, and projection distance loss aligning the visual space and semantic spaces of old seen classes. Experiments demonstrate the efficacy of our model in detecting objects from both the seen and unseen classes, outperforming the alternative models on Pascal VOC and MSCOCO datasets.

Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers

Most recent semantic segmentation methods adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract/semantic visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated/atrous convolutions or inserting attention modules. However, the encoder-decoder based FCN architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating semantic segmentation as a sequence-to-sequence prediction task. Specifically, we deploy a pure transformer (ie, without convolution and resolution reduction) to encode an image as a sequence of patches. With the global context modeled in every layer of the transformer, this encoder can be combined with a simple decoder to provide a powerful segmentation model, termed SEgmentation TRansformer (SETR). Extensive experiments show that SETR achieves new state of the art on ADE20K (50.28% mIoU), Pascal Context (55.83% mIoU) and competitive results on Cityscapes. Particularly, we achieve the first (44.42% mIoU) position in the highly competitive ADE20K test server leaderboard.