Convolution has been the core ingredient of modern neural networks, triggering the surge of deep learning in vision. In this work, we rethink the inherent principles of standard convolution for vision tasks, specifically spatial-agnostic and channel-specific. Instead, we present a novel atomic operation for deep neural networks by inverting the aforementioned design principles of convolution, coined as involution. We additionally demystify the recent popular self-attention operator and subsume it into our involution family as an over-complicated instantiation. The proposed involution operator could be leveraged as fundamental bricks to build the new generation of neural networks for visual recognition, powering different deep learning models on several prevalent benchmarks, including ImageNet classification, COCO detection and segmentation, together with Cityscapes segmentation. Our involution-based models improve the performance of convolutional baselines using ResNet-50 by up to 1.6% top-1 accuracy, 2.5% and 2.4% bounding box AP, and 4.7% mean IoU absolutely while compressing the computational cost to 66%, 65%, 72%, and 57% on the above benchmarks, respectively. Code and pre-trained models for all the tasks are available at https://github.com/d-li14/involution.
Fine-grained image recognition is very challenging due to the difficulty of capturing both semantic global features and discriminative local features. Meanwhile, these two features are not easy to be integrated, which are even conflicting when used simultaneously. In this paper, a retrieval-based coarse-to-fine framework is proposed, where we re-rank the TopN classification results by using the local region enhanced embedding features to improve the Top1 accuracy (based on the observation that the correct category usually resides in TopN results). To obtain the discriminative regions for distinguishing the fine-grained images, we introduce a weakly-supervised method to train a box generating branch with only image-level labels. In addition, to learn more effective semantic global features, we design a multi-level loss over an automatically constructed hierarchical category structure. Experimental results show that our method achieves state-of-the-art performance on three benchmarks: CUB-200-2011, Stanford Cars, and FGVC Aircraft. Also, visualizations and analysis are provided for better understanding.
Most typical click models assume that the probability of a document to be examined by users only depends on position, such as PBM and UBM. It works well in various kinds of search engines. However, in a search engine where massive candidate documents display images as responses to the query, the examination probability should not only depend on position. The visual appearance of an image-oriented document also plays an important role in its opportunity to be examined. In this paper, we assume that vision bias exists in an image-oriented search engine as another crucial factor affecting the examination probability aside from position. Specifically, we apply this assumption to classical click models and propose an extended model, to better capture the examination probabilities of documents. We use regression-based EM algorithm to predict the vision bias given the visual features extracted from candidate documents. Empirically, we evaluate our model on a dataset developed from a real-world online image-oriented search engine, and demonstrate that our proposed model can achieve significant improvements over its baseline model in data fitness and sparsity handling.
Multiple-object tracking(MOT) is mostly dominated by complex and multi-step tracking-by-detection algorithm, which performs object detection, feature extraction and temporal association, separately. Query-key mechanism in single-object tracking(SOT), which tracks the object of the current frame by object feature of the previous frame, has great potential to set up a simple joint-detection-and-tracking MOT paradigm. Nonetheless, the query-key method is seldom studied due to its inability to detect new-coming objects. In this work, we propose TransTrack, a baseline for MOT with Transformer. It takes advantage of query-key mechanism and introduces a set of learned object queries into the pipeline to enable detecting new-coming objects. TransTrack has three main advantages: (1) It is an online joint-detection-and-tracking pipeline based on query-key mechanism. Complex and multi-step components in the previous methods are simplified. (2) It is a brand new architecture based on Transformer. The learned object query detects objects in the current frame. The object feature query from the previous frame associates those current objects with the previous ones. (3) For the first time, we demonstrate a much simple and effective method based on query-key mechanism and Transformer architecture could achieve competitive 65.8\% MOTA on the MOT17 challenge dataset. We hope TransTrack can provide a new perspective for multiple-object tracking. The code is available at: \url{https://github.com/PeizeSun/TransTrack}.
Generative adversarial networks (GANs) have achieved remarkable progress in recent years, but the continuously growing scale of models make them challenging to deploy widely in practical applications. In particular, for real-time generation tasks, different devices require generators of different sizes due to varying computing power. In this paper, we introduce slimmable GANs (SlimGANs), which can flexibly switch the width of the generator to accommodate various quality-efficiency trade-offs at runtime. Specifically, we leverage multiple discriminators that share partial parameters to train the slimmable generator. To facilitate the consistency between generators of different widths, we present a stepwise inplace distillation technique that encourages narrow generators to learn from wide ones. As for class-conditional generation, we propose a sliceable conditional batch normalization that incorporates the label information into different widths. Our methods are validated, both quantitatively and qualitatively, by extensive experiments and a detailed ablation study.
End-to-end one-stage object detection trailed thus far. This paper discovers that the lack of classification cost between sample and ground-truth in label assignment is the main obstacle for one-stage detectors to remove Non-maximum Suppression(NMS) and reach end-to-end. Existing one-stage object detectors assign labels by only location cost, e.g. box IoU or point distance. Without classification cost, sole location cost leads to redundant boxes of high confidence scores in inference, making NMS necessary post-processing. To design an end-to-end one-stage object detector, we propose Minimum Cost Assignment. The cost is the summation of classification cost and location cost between sample and ground-truth. For each object ground-truth, only one sample of minimum cost is assigned as the positive sample; others are all negative samples. To evaluate the effectiveness of our method, we design an extremely simple one-stage detector named OneNet. Our results show that when trained with Minimum Cost Assignment, OneNet avoids producing duplicated boxes and achieves to end-to-end detector. On COCO dataset, OneNet achieves 35.0 AP/80 FPS and 37.7 AP/50 FPS with image size of 512 pixels. We hope OneNet could serve as an effective baseline for end-to-end one-stage object detection. The code is available at: \url{https://github.com/PeizeSun/OneNet}.
Although deep learning based methods have achieved great progress in unsupervised video object segmentation, difficult scenarios (e.g., visual similarity, occlusions, and appearance changing) are still not well-handled. To alleviate these issues, we propose a novel Focus on Foreground Network (F2Net), which delves into the intra-inter frame details for the foreground objects and thus effectively improve the segmentation performance. Specifically, our proposed network consists of three main parts: Siamese Encoder Module, Center Guiding Appearance Diffusion Module, and Dynamic Information Fusion Module. Firstly, we take a siamese encoder to extract the feature representations of paired frames (reference frame and current frame). Then, a Center Guiding Appearance Diffusion Module is designed to capture the inter-frame feature (dense correspondences between reference frame and current frame), intra-frame feature (dense correspondences in current frame), and original semantic feature of current frame. Specifically, we establish a Center Prediction Branch to predict the center location of the foreground object in current frame and leverage the center point information as spatial guidance prior to enhance the inter-frame and intra-frame feature extraction, and thus the feature representation considerably focus on the foreground objects. Finally, we propose a Dynamic Information Fusion Module to automatically select relatively important features through three aforementioned different level features. Extensive experiments on DAVIS2016, Youtube-object, and FBMS datasets show that our proposed F2Net achieves the state-of-the-art performance with significant improvement.
We present Sparse R-CNN, a purely sparse method for object detection in images. Existing works on object detection heavily rely on dense object candidates, such as $k$ anchor boxes pre-defined on all grids of image feature map of size $H\times W$. In our method, however, a fixed sparse set of learned object proposals, total length of $N$, are provided to object recognition head to perform classification and location. By eliminating $HWk$ (up to hundreds of thousands) hand-designed object candidates to $N$ (e.g. 100) learnable proposals, Sparse R-CNN completely avoids all efforts related to object candidates design and many-to-one label assignment. More importantly, final predictions are directly output without non-maximum suppression post-procedure. Sparse R-CNN demonstrates accuracy, run-time and training convergence performance on par with the well-established detector baselines on the challenging COCO dataset, e.g., achieving 44.5 AP in standard $3\times$ training schedule and running at 22 fps using ResNet-50 FPN model. We hope our work could inspire re-thinking the convention of dense prior in object detectors. The code is available at: https://github.com/PeizeSun/SparseR-CNN.
Visual Semantic Embedding (VSE) is a dominant approach for vision-language retrieval, which aims at learning a deep embedding space such that visual data are embedded close to their semantic text labels or descriptions. Recent VSE models use complex methods to better contextualize and aggregate multi-modal features into holistic embeddings. However, we discover that surprisingly simple (but carefully selected) global pooling functions (e.g., max pooling) outperform those complex models, across different feature extractors. Despite its simplicity and effectiveness, seeking the best pooling function for different data modality and feature extractor is costly and tedious, especially when the size of features varies (e.g., text, video). Therefore, we propose a Generalized Pooling Operator (GPO), which learns to automatically adapt itself to the best pooling strategy for different features, requiring no manual tuning while staying effective and efficient. We extend the VSE model using this proposed GPO and denote it as VSE$\infty$. Without bells and whistles, VSE$\infty$ outperforms previous VSE methods significantly on image-text retrieval benchmarks across popular feature extractors. With a simple adaptation, variants of VSE$\infty$ further demonstrate its strength by achieving the new state of the art on two video-text retrieval datasets. Comprehensive experiments and visualizations confirm that GPO always discovers the best pooling strategy and can be a plug-and-play feature aggregation module for standard VSE models.