Recent research that applies Transformer-based architectures to image captioning has resulted in state-of-the-art image captioning performance, capitalising on the success of Transformers on natural language tasks. Unfortunately, though these models work well, one major flaw is their large model sizes. To this end, we present three parameter reduction methods for image captioning Transformers: Radix Encoding, cross-layer parameter sharing, and attention parameter sharing. By combining these methods, our proposed ACORT models have 3.7x to 21.6x fewer parameters than the baseline model without compromising test performance. Results on the MS-COCO dataset demonstrate that our ACORT models are competitive against baselines and SOTA approaches, with CIDEr score >=126. Finally, we present qualitative results and ablation studies to demonstrate the efficacy of the proposed changes further. Code and pre-trained models are publicly available at https://github.com/jiahuei/sparse-image-captioning.
Generative Adversarial Networks (GANs) have been widely applied in modeling diverse image distributions. However, despite its impressive applications, the structure of the latent space in GANs largely remains as a black-box, leaving its controllable generation an open problem, especially when spurious correlations between different semantic attributes exist in the image distributions. To address this problem, previous methods typically learn linear directions or individual channels that control semantic attributes in the image space. However, they often suffer from imperfect disentanglement, or are unable to obtain multi-directional controls. In this work, in light of the above challenges, we propose a novel approach that discovers nonlinear controls, which enables multi-directional manipulation as well as effective disentanglement, based on gradient information in the learned GAN latent space. More specifically, we first learn interpolation directions by following the gradients from classification networks trained separately on the attributes, and then navigate the latent space by exclusively controlling channels activated for the target attribute in the learned directions. Empirically, with small training data, our approach is able to gain fine-grained controls over a diverse set of bi-directional and multi-directional attributes, and we showcase its ability to achieve disentanglement significantly better than state-of-the-art methods both qualitatively and quantitatively.
Few-shot classification which aims to recognize unseen classes using very limited samples has attracted more and more attention. Usually, it is formulated as a metric learning problem. The core issue of few-shot classification is how to learn (1) consistent representations for images in both support and query sets and (2) effective metric learning for images between support and query sets. In this paper, we show that the two challenges can be well modeled simultaneously via a unified Query-Support TransFormer (QSFormer) model. To be specific,the proposed QSFormer involves global query-support sample Transformer (sampleFormer) branch and local patch Transformer (patchFormer) learning branch. sampleFormer aims to capture the dependence of samples in support and query sets for image representation. It adopts the Encoder, Decoder and Cross-Attention to respectively model the Support, Query (image) representation and Metric learning for few-shot classification task. Also, as a complementary to global learning branch, we adopt a local patch Transformer to extract structural representation for each image sample by capturing the long-range dependence of local image patches. In addition, a novel Cross-scale Interactive Feature Extractor (CIFE) is proposed to extract and fuse multi-scale CNN features as an effective backbone module for the proposed few-shot learning method. All modules are integrated into a unified framework and trained in an end-to-end manner. Extensive experiments on four popular datasets demonstrate the effectiveness and superiority of the proposed QSFormer.
Reference-guided image inpainting restores image pixels by leveraging the content from another reference image. The previous state-of-the-art, TransFill, warps the source image with multiple homographies, and fuses them together for hole filling. Inspired by structure from motion pipelines and recent progress in monocular depth estimation, we propose a more principled approach that does not require heuristic planar assumptions. We leverage a monocular depth estimate and predict relative pose between cameras, then align the reference image to the target by a differentiable 3D reprojection and a joint optimization of relative pose and depth map scale and offset. Our approach achieves state-of-the-art performance on both RealEstate10K and MannequinChallenge dataset with large baselines, complex geometry and extreme camera motions. We experimentally verify our approach is also better at handling large holes.
Contrastive learning, a self-supervised learning method that can learn representations from unlabeled data, has been developed promisingly. Many methods of contrastive learning depend on data augmentation techniques, which generate different views from the original signal. However, tuning policies and hyper-parameters for more effective data augmentation methods in contrastive learning is often time and resource-consuming. Researchers have designed approaches to automatically generate new views for some input signals, especially on the image data. But the view-learning method is not well developed for time-series data. In this work, we propose a simple but effective module for automating view generation for time-series data in contrastive learning, named learning views for time-series data (LEAVES). The proposed module learns the hyper-parameters for augmentations using adversarial training in contrastive learning. We validate the effectiveness of the proposed method using multiple time-series datasets. The experiments demonstrate that the proposed method is more effective in finding reasonable views and performs downstream tasks better than the baselines, including manually tuned augmentation-based contrastive learning methods and SOTA methods.
Deep learning models dealing with image understanding in real-world settings must be able to adapt to a wide variety of tasks across different domains. Domain adaptation and class incremental learning deal with domain and task variability separately, whereas their unified solution is still an open problem. We tackle both facets of the problem together, taking into account the semantic shift within both input and label spaces. We start by formally introducing continual learning under task and domain shift. Then, we address the proposed setup by using style transfer techniques to extend knowledge across domains when learning incremental tasks and a robust distillation framework to effectively recollect task knowledge under incremental domain shift. The devised framework (LwS, Learning with Style) is able to generalize incrementally acquired task knowledge across all the domains encountered, proving to be robust against catastrophic forgetting. Extensive experimental evaluation on multiple autonomous driving datasets shows how the proposed method outperforms existing approaches, which prove to be ill-equipped to deal with continual semantic segmentation under both task and domain shift.
Segmentation, i.e., the partitioning of volumetric data into components, is a crucial task in many image processing applications ever since such data could be generated. Most existing applications nowadays, specifically CNNs, make use of voxelwise classification systems which need to be trained on a large number of annotated training volumes. However, in many practical applications such data sets are seldom available and the generation of annotations is time-consuming and cumbersome. In this paper, we introduce a novel voxelwise segmentation method based on active learning on geometric features. Our method uses interactively provided seed points to train a voxelwise classifier based entirely on local information. The combination of an ad hoc incorporation of domain knowledge and local processing results in a flexible yet efficient segmentation method that is applicable to three-dimensional volumes without size restrictions. We illustrate the potential and flexibility of our approach by applying it to selected computed tomography scans where we perform different segmentation tasks to scans from different domains and of different sizes.
Multimodal transformer exhibits high capacity and flexibility to align image and text for visual grounding. However, the encoder-only grounding framework (e.g., TransVG) suffers from heavy computation due to the self-attention operation with quadratic time complexity. To address this issue, we present a new multimodal transformer architecture, coined as Dynamic MDETR, by decoupling the whole grounding process into encoding and decoding phases. The key observation is that there exists high spatial redundancy in images. Thus, we devise a new dynamic multimodal transformer decoder by exploiting this sparsity prior to speed up the visual grounding process. Specifically, our dynamic decoder is composed of a 2D adaptive sampling module and a text-guided decoding module. The sampling module aims to select these informative patches by predicting the offsets with respect to a reference point, while the decoding module works for extracting the grounded object information by performing cross attention between image features and text features. These two modules are stacked alternatively to gradually bridge the modality gap and iteratively refine the reference point of grounded object, eventually realizing the objective of visual grounding. Extensive experiments on five benchmarks demonstrate that our proposed Dynamic MDETR achieves competitive trade-offs between computation and accuracy. Notably, using only 9% feature points in the decoder, we can reduce ~44% GLOPs of the multimodal transformer, but still get higher accuracy than the encoder-only counterpart. In addition, to verify its generalization ability and scale up our Dynamic MDETR, we build the first one-stage CLIP empowered visual grounding framework, and achieve the state-of-the-art performance on these benchmarks.
Dual encoders and cross encoders have been widely used for image-text retrieval. Between the two, the dual encoder encodes the image and text independently followed by a dot product, while the cross encoder jointly feeds image and text as the input and performs dense multi-modal fusion. These two architectures are typically modeled separately without interaction. In this work, we propose LoopITR, which combines them in the same network for joint learning. Specifically, we let the dual encoder provide hard negatives to the cross encoder, and use the more discriminative cross encoder to distill its predictions back to the dual encoder. Both steps are efficiently performed together in the same model. Our work centers on empirical analyses of this combined architecture, putting the main focus on the design of the distillation objective. Our experimental results highlight the benefits of training the two encoders in the same network, and demonstrate that distillation can be quite effective with just a few hard negative examples. Experiments on two standard datasets (Flickr30K and COCO) show our approach achieves state-of-the-art dual encoder performance when compared with approaches using a similar amount of data.
Knowledge distillation (KD) is an effective tool for compressing deep classification models for edge devices. However, the performance of KD is affected by the large capacity gap between the teacher and student networks. Recent methods have resorted to a multiple teacher assistant (TA) setting for KD, which sequentially decreases the size of the teacher model to relatively bridge the size gap between these models. This paper proposes a new technique called Curriculum Expert Selection for Knowledge Distillation (CES-KD) to efficiently enhance the learning of a compact student under the capacity gap problem. This technique is built upon the hypothesis that a student network should be guided gradually using stratified teaching curriculum as it learns easy (hard) data samples better and faster from a lower (higher) capacity teacher network. Specifically, our method is a gradual TA-based KD technique that selects a single teacher per input image based on a curriculum driven by the difficulty in classifying the image. In this work, we empirically verify our hypothesis and rigorously experiment with CIFAR-10, CIFAR-100, CINIC-10, and ImageNet datasets and show improved accuracy on VGG-like models, ResNets, and WideResNets architectures.