Freezing the pre-trained backbone has become a standard paradigm to avoid overfitting in few-shot segmentation. In this paper, we rethink the paradigm and explore a new regime: {\em fine-tuning a small part of parameters in the backbone}. We present a solution to overcome the overfitting problem, leading to better model generalization on learning novel classes. Our method decomposes backbone parameters into three successive matrices via the Singular Value Decomposition (SVD), then {\em only fine-tunes the singular values} and keeps others frozen. The above design allows the model to adjust feature representations on novel classes while maintaining semantic clues within the pre-trained backbone. We evaluate our {\em Singular Value Fine-tuning (SVF)} approach on various few-shot segmentation methods with different backbones. We achieve state-of-the-art results on both Pascal-5$^i$ and COCO-20$^i$ across 1-shot and 5-shot settings. Hopefully, this simple baseline will encourage researchers to rethink the role of backbone fine-tuning in few-shot settings. The source code and models will be available at \url{https://github.com/syp2ysy/SVF}.
In this paper, we present a model pretraining technique, named MaskOCR, for text recognition. Our text recognition architecture is an encoder-decoder transformer: the encoder extracts the patch-level representations, and the decoder recognizes the text from the representations. Our approach pretrains both the encoder and the decoder in a sequential manner. (i) We pretrain the encoder in a self-supervised manner over a large set of unlabeled real text images. We adopt the masked image modeling approach, which shows the effectiveness for general images, expecting that the representations take on semantics. (ii) We pretrain the decoder over a large set of synthesized text images in a supervised manner and enhance the language modeling capability of the decoder by randomly masking some text image patches occupied by characters input to the encoder and accordingly the representations input to the decoder. Experiments show that the proposed MaskOCR approach achieves superior results on the benchmark datasets, including Chinese and English text images.
Few-shot font generation (FFG), which aims to generate a new font with a few examples, is gaining increasing attention due to the significant reduction in labor cost. A typical FFG pipeline considers characters in a standard font library as content glyphs and transfers them to a new target font by extracting style information from the reference glyphs. Most existing solutions explicitly disentangle content and style of reference glyphs globally or component-wisely. However, the style of glyphs mainly lies in the local details, i.e. the styles of radicals, components, and strokes together depict the style of a glyph. Therefore, even a single character can contain different styles distributed over spatial locations. In this paper, we propose a new font generation approach by learning 1) the fine-grained local styles from references, and 2) the spatial correspondence between the content and reference glyphs. Therefore, each spatial location in the content glyph can be assigned with the right fine-grained style. To this end, we adopt cross-attention over the representation of the content glyphs as the queries and the representations of the reference glyphs as the keys and values. Instead of explicitly disentangling global or component-wise modeling, the cross-attention mechanism can attend to the right local styles in the reference glyphs and aggregate the reference styles into a fine-grained style representation for the given content glyphs. The experiments show that the proposed method outperforms the state-of-the-art methods in FFG. In particular, the user studies also demonstrate the style consistency of our approach significantly outperforms previous methods.
Despite the broad range of algorithms for Approximate Nearest Neighbor Search, most empirical evaluations of algorithms have focused on smaller datasets, typically of 1 million points~\citep{Benchmark}. However, deploying recent advances in embedding based techniques for search, recommendation and ranking at scale require ANNS indices at billion, trillion or larger scale. Barring a few recent papers, there is limited consensus on which algorithms are effective at this scale vis-\`a-vis their hardware cost. This competition compares ANNS algorithms at billion-scale by hardware cost, accuracy and performance. We set up an open source evaluation framework and leaderboards for both standardized and specialized hardware. The competition involves three tracks. The standard hardware track T1 evaluates algorithms on an Azure VM with limited DRAM, often the bottleneck in serving billion-scale indices, where the embedding data can be hundreds of GigaBytes in size. It uses FAISS~\citep{Faiss17} as the baseline. The standard hardware track T2 additional allows inexpensive SSDs in addition to the limited DRAM and uses DiskANN~\citep{DiskANN19} as the baseline. The specialized hardware track T3 allows any hardware configuration, and again uses FAISS as the baseline. We compiled six diverse billion-scale datasets, four newly released for this competition, that span a variety of modalities, data types, dimensions, deep learning models, distance functions and sources. The outcome of the competition was ranked leaderboards of algorithms in each track based on recall at a query throughput threshold. Additionally, for track T3, separate leaderboards were created based on recall as well as cost-normalized and power-normalized query throughput.
The head swapping task aims at flawlessly placing a source head onto a target body, which is of great importance to various entertainment scenarios. While face swapping has drawn much attention, the task of head swapping has rarely been explored, particularly under the few-shot setting. It is inherently challenging due to its unique needs in head modeling and background blending. In this paper, we present the Head Swapper (HeSer), which achieves few-shot head swapping in the wild through two delicately designed modules. Firstly, a Head2Head Aligner is devised to holistically migrate pose and expression information from the target to the source head by examining multi-scale information. Secondly, to tackle the challenges of skin color variations and head-background mismatches in the swapping procedure, a Head2Scene Blender is introduced to simultaneously modify facial skin color and fill mismatched gaps in the background around the head. Particularly, seamless blending is achieved with the help of a Semantic-Guided Color Reference Creation procedure and a Blending UNet. Extensive experiments demonstrate that the proposed method produces superior head swapping results in a variety of scenes.
Human-Object Interaction Detection tackles the problem of joint localization and classification of human object interactions. Existing HOI transformers either adopt a single decoder for triplet prediction, or utilize two parallel decoders to detect individual objects and interactions separately, and compose triplets by a matching process. In contrast, we decouple the triplet prediction into human-object pair detection and interaction classification. Our main motivation is that detecting the human-object instances and classifying interactions accurately needs to learn representations that focus on different regions. To this end, we present Disentangled Transformer, where both encoder and decoder are disentangled to facilitate learning of two sub-tasks. To associate the predictions of disentangled decoders, we first generate a unified representation for HOI triplets with a base decoder, and then utilize it as input feature of each disentangled decoder. Extensive experiments show that our method outperforms prior work on two public HOI benchmarks by a sizeable margin. Code will be available.
Birds-eye-view (BEV) semantic segmentation is critical for autonomous driving for its powerful spatial representation ability. It is challenging to estimate the BEV semantic maps from monocular images due to the spatial gap, since it is implicitly required to realize both the perspective-to-BEV transformation and segmentation. We present a novel two-stage Geometry Prior-based Transformation framework named GitNet, consisting of (i) the geometry-guided pre-alignment and (ii) ray-based transformer. In the first stage, we decouple the BEV segmentation into the perspective image segmentation and geometric prior-based mapping, with explicit supervision by projecting the BEV semantic labels onto the image plane to learn visibility-aware features and learnable geometry to translate into BEV space. Second, the pre-aligned coarse BEV features are further deformed by ray-based transformers to take visibility knowledge into account. GitNet achieves the leading performance on the challenging nuScenes and Argoverse Datasets. The code will be publicly available.
Most existing unsupervised person re-identification (Re-ID) methods use clustering to generate pseudo labels for model training. Unfortunately, clustering sometimes mixes different true identities together or splits the same identity into two or more sub clusters. Training on these noisy clusters substantially hampers the Re-ID accuracy. Due to the limited samples in each identity, we suppose there may lack some underlying information to well reveal the accurate clusters. To discover these information, we propose an Implicit Sample Extension (\OurWholeMethod) method to generate what we call support samples around the cluster boundaries. Specifically, we generate support samples from actual samples and their neighbouring clusters in the embedding space through a progressive linear interpolation (PLI) strategy. PLI controls the generation with two critical factors, i.e., 1) the direction from the actual sample towards its K-nearest clusters and 2) the degree for mixing up the context information from the K-nearest clusters. Meanwhile, given the support samples, ISE further uses a label-preserving loss to pull them towards their corresponding actual samples, so as to compact each cluster. Consequently, ISE reduces the "sub and mixed" clustering errors, thus improving the Re-ID performance. Extensive experiments demonstrate that the proposed method is effective and achieves state-of-the-art performance for unsupervised person Re-ID. Code is available at: \url{https://github.com/PaddlePaddle/PaddleClas}.
While local-window self-attention performs notably in vision tasks, it suffers from limited receptive field and weak modeling capability issues. This is mainly because it performs self-attention within non-overlapped windows and shares weights on the channel dimension. We propose MixFormer to find a solution. First, we combine local-window self-attention with depth-wise convolution in a parallel design, modeling cross-window connections to enlarge the receptive fields. Second, we propose bi-directional interactions across branches to provide complementary clues in the channel and spatial dimensions. These two designs are integrated to achieve efficient feature mixing among windows and dimensions. Our MixFormer provides competitive results on image classification with EfficientNet and shows better results than RegNet and Swin Transformer. Performance in downstream tasks outperforms its alternatives by significant margins with less computational costs in 5 dense prediction tasks on MS COCO, ADE20k, and LVIS. Code is available at \url{https://github.com/PaddlePaddle/PaddleClas}.
In this work, we introduce Dual Attention Vision Transformers (DaViT), a simple yet effective vision transformer architecture that is able to capture global context while maintaining computational efficiency. We propose approaching the problem from an orthogonal angle: exploiting self-attention mechanisms with both "spatial tokens" and "channel tokens". With spatial tokens, the spatial dimension defines the token scope, and the channel dimension defines the token feature dimension. With channel tokens, we have the inverse: the channel dimension defines the token scope, and the spatial dimension defines the token feature dimension. We further group tokens along the sequence direction for both spatial and channel tokens to maintain the linear complexity of the entire model. We show that these two self-attentions complement each other: (i) since each channel token contains an abstract representation of the entire image, the channel attention naturally captures global interactions and representations by taking all spatial positions into account when computing attention scores between channels; (ii) the spatial attention refines the local representations by performing fine-grained interactions across spatial locations, which in turn helps the global information modeling in channel attention. Extensive experiments show our DaViT achieves state-of-the-art performance on four different tasks with efficient computations. Without extra data, DaViT-Tiny, DaViT-Small, and DaViT-Base achieve 82.8%, 84.2%, and 84.6% top-1 accuracy on ImageNet-1K with 28.3M, 49.7M, and 87.9M parameters, respectively. When we further scale up DaViT with 1.5B weakly supervised image and text pairs, DaViT-Gaint reaches 90.4% top-1 accuracy on ImageNet-1K. Code is available at https://github.com/dingmyu/davit.