知识蒸馏在分类中取得了巨大的成功，但是，仍然有挑战性。在用于检测的典型图像中，来自不同位置的表示可能对检测目标具有不同的贡献，使蒸馏难以平衡。在本文中，我们提出了一种有条件的蒸馏框架来蒸馏出所需的知识，即关于每个例子的分类和本地化有益的知识。该框架引入了一种可学习的条件解码模块，其将每个目标实例检索为查询的信息。具体而言，我们将条件信息编码为查询并使用教师的表示作为键。查询和键之间的注意用于测量不同特征的贡献，由本地化识别敏感辅助任务指导。广泛的实验表明了我们的方法的功效：我们在各种环境下观察到令人印象深刻的改进。值得注意的是，在1倍计划下，我们将通过37.4至40.7地图（+3.3）与Reset-50骨架的Restinetet提升。代码已在https://github.com/megvii-research/icd上发布。

在本文中，我们提出了一种用于一般物体检测的第一自蒸馏框架，称为LGD（标签引导自蒸馏）。以前的研究依赖于强大的预酝酿教师，以提供在现实世界方案中可能无法使用的指导知识。相反，我们通过对象之间的关系间和帧间关系建模来生成一个有效的知识，只需要学生表示和常规标签。具体而言，我们的框架涉及稀疏的标签外观编码，对象间关系适应和对象内的知识映射，以获得指导知识。他们在培训阶段共同形成隐式教师，动态依赖标签和不断发展的学生表示。 LGD中的模块与学生检测器的端到端训练，并在推理中丢弃。实验上，LGD在各种探测器，数据集和广泛的任务上获得了体面的结果，如实例分段。例如，在MS-Coco DataSet中，LGD将Reset-50下的REDINENT改善2倍单尺度培训，从36.2％到39.0％地图（+ 2.8％）。它在2倍多尺度培训下使用Resnext-101 DCN V2等FCO的探测器增加了更强大的探测器，从46.1％到47.9％（+ 1.8％）。与古典教师的方法FGFI相比，LGD不仅在不需要佩金的教师而且还可以降低固有的学生学习超出51％的培训成本。

An unbiased scene graph generation (SGG) algorithm referred to as Skew Class-balanced Re-weighting (SCR) is proposed for considering the unbiased predicate prediction caused by the long-tailed distribution. The prior works focus mainly on alleviating the deteriorating performances of the minority predicate predictions, showing drastic dropping recall scores, i.e., losing the majority predicate performances. It has not yet correctly analyzed the trade-off between majority and minority predicate performances in the limited SGG datasets. In this paper, to alleviate the issue, the Skew Class-balanced Re-weighting (SCR) loss function is considered for the unbiased SGG models. Leveraged by the skewness of biased predicate predictions, the SCR estimates the target predicate weight coefficient and then re-weights more to the biased predicates for better trading-off between the majority predicates and the minority ones. Extensive experiments conducted on the standard Visual Genome dataset and Open Image V4 \& V6 show the performances and generality of the SCR with the traditional SGG models.

In the field of cross-modal retrieval, single encoder models tend to perform better than dual encoder models, but they suffer from high latency and low throughput. In this paper, we present a dual encoder model called BagFormer that utilizes a cross modal interaction mechanism to improve recall performance without sacrificing latency and throughput. BagFormer achieves this through the use of bag-wise interactions, which allow for the transformation of text to a more appropriate granularity and the incorporation of entity knowledge into the model. Our experiments demonstrate that BagFormer is able to achieve results comparable to state-of-the-art single encoder models in cross-modal retrieval tasks, while also offering efficient training and inference with 20.72 times lower latency and 25.74 times higher throughput.

Deep learning has been widely used for protein engineering. However, it is limited by the lack of sufficient experimental data to train an accurate model for predicting the functional fitness of high-order mutants. Here, we develop SESNet, a supervised deep-learning model to predict the fitness for protein mutants by leveraging both sequence and structure information, and exploiting attention mechanism. Our model integrates local evolutionary context from homologous sequences, the global evolutionary context encoding rich semantic from the universal protein sequence space and the structure information accounting for the microenvironment around each residue in a protein. We show that SESNet outperforms state-of-the-art models for predicting the sequence-function relationship on 26 deep mutational scanning datasets. More importantly, we propose a data augmentation strategy by leveraging the data from unsupervised models to pre-train our model. After that, our model can achieve strikingly high accuracy in prediction of the fitness of protein mutants, especially for the higher order variants (> 4 mutation sites), when finetuned by using only a small number of experimental mutation data (<50). The strategy proposed is of great practical value as the required experimental effort, i.e., producing a few tens of experimental mutation data on a given protein, is generally affordable by an ordinary biochemical group and can be applied on almost any protein.

Three-dimensional (3D) ultrasound imaging technique has been applied for scoliosis assessment, but current assessment method only uses coronal projection image and cannot illustrate the 3D deformity and vertebra rotation. The vertebra detection is essential to reveal 3D spine information, but the detection task is challenging due to complex data and limited annotations. We propose VertMatch, a two-step framework to detect vertebral structures in 3D ultrasound volume by utilizing unlabeled data in semi-supervised manner. The first step is to detect the possible positions of structures on transverse slice globally, and then the local patches are cropped based on detected positions. The second step is to distinguish whether the patches contain real vertebral structures and screen the predicted positions from the first step. VertMatch develops three novel components for semi-supervised learning: for position detection in the first step, (1) anatomical prior is used to screen pseudo labels generated from confidence threshold method; (2) multi-slice consistency is used to utilize more unlabeled data by inputting multiple adjacent slices; (3) for patch identification in the second step, the categories are rebalanced in each batch to solve imbalance problem. Experimental results demonstrate that VertMatch can detect vertebra accurately in ultrasound volume and outperforms state-of-the-art methods. VertMatch is also validated in clinical application on forty ultrasound scans, and it can be a promising approach for 3D assessment of scoliosis.

Image captioning is one of the straightforward tasks that can take advantage of large-scale web-crawled data which provides rich knowledge about the visual world for a captioning model. However, since web-crawled data contains image-text pairs that are aligned at different levels, the inherent noises (e.g., misaligned pairs) make it difficult to learn a precise captioning model. While the filtering strategy can effectively remove noisy data, however, it leads to a decrease in learnable knowledge and sometimes brings about a new problem of data deficiency. To take the best of both worlds, we propose a noise-aware learning framework, which learns rich knowledge from the whole web-crawled data while being less affected by the noises. This is achieved by the proposed quality controllable model, which is learned using alignment levels of the image-text pairs as an additional control signal during training. The alignment-conditioned training allows the model to generate high-quality captions of well-aligned by simply setting the control signal to desired alignment level at inference time. Through in-depth analysis, we show that our controllable captioning model is effective in handling noise. In addition, with two tasks of zero-shot captioning and text-to-image retrieval using generated captions (i.e., self-retrieval), we also demonstrate our model can produce high-quality captions in terms of descriptiveness and distinctiveness. Code is available at \url{https://github.com/kakaobrain/noc}.

Diffusion Probabilistic Models (DPMs) have shown a powerful capacity of generating high-quality image samples. Recently, diffusion autoencoders (Diff-AE) have been proposed to explore DPMs for representation learning via autoencoding. Their key idea is to jointly train an encoder for discovering meaningful representations from images and a conditional DPM as the decoder for reconstructing images. Considering that training DPMs from scratch will take a long time and there have existed numerous pre-trained DPMs, we propose \textbf{P}re-trained \textbf{D}PM \textbf{A}uto\textbf{E}ncoding (\textbf{PDAE}), a general method to adapt existing pre-trained DPMs to the decoders for image reconstruction, with better training efficiency and performance than Diff-AE. Specifically, we find that the reason that pre-trained DPMs fail to reconstruct an image from its latent variables is due to the information loss of forward process, which causes a gap between their predicted posterior mean and the true one. From this perspective, the classifier-guided sampling method can be explained as computing an extra mean shift to fill the gap, reconstructing the lost class information in samples. These imply that the gap corresponds to the lost information of the image, and we can reconstruct the image by filling the gap. Drawing inspiration from this, we employ a trainable model to predict a mean shift according to encoded representation and train it to fill as much gap as possible, in this way, the encoder is forced to learn as much information as possible from images to help the filling. By reusing a part of network of pre-trained DPMs and redesigning the weighting scheme of diffusion loss, PDAE can learn meaningful representations from images efficiently. Extensive experiments demonstrate the effectiveness, efficiency and flexibility of PDAE.

Automatic image colorization is a particularly challenging problem. Due to the high illness of the problem and multi-modal uncertainty, directly training a deep neural network usually leads to incorrect semantic colors and low color richness. Existing transformer-based methods can deliver better results but highly depend on hand-crafted dataset-level empirical distribution priors. In this work, we propose DDColor, a new end-to-end method with dual decoders, for image colorization. More specifically, we design a multi-scale image decoder and a transformer-based color decoder. The former manages to restore the spatial resolution of the image, while the latter establishes the correlation between semantic representations and color queries via cross-attention. The two decoders incorporate to learn semantic-aware color embedding by leveraging the multi-scale visual features. With the help of these two decoders, our method succeeds in producing semantically consistent and visually plausible colorization results without any additional priors. In addition, a simple but effective colorfulness loss is introduced to further improve the color richness of generated results. Our extensive experiments demonstrate that the proposed DDColor achieves significantly superior performance to existing state-of-the-art works both quantitatively and qualitatively. Codes will be made publicly available.

Obtaining ground truth data in medical imaging has difficulties due to the fact that it requires a lot of annotating time from the experts in the field. Also, when trained with supervised learning, it detects only the cases included in the labels. In real practice, we want to also open to other possibilities than the named cases while examining the medical images. As a solution, the need for anomaly detection that can detect and localize abnormalities by learning the normal characteristics using only normal images is emerging. With medical image data, we can design either 2D or 3D networks of self-supervised learning for anomaly detection task. Although 3D networks, which learns 3D structures of the human body, show good performance in 3D medical image anomaly detection, they cannot be stacked in deeper layers due to memory problems. While 2D networks have advantage in feature detection, they lack 3D context information. In this paper, we develop a method for combining the strength of the 3D network and the strength of the 2D network through joint embedding. We also propose the pretask of self-supervised learning to make it possible for the networks to learn efficiently. Through the experiments, we show that the proposed method achieves better performance in both classification and segmentation tasks compared to the SoTA method.