BIFRÖST: 3D-Aware Image compositing with Language Instructions

This paper introduces Bifr\"ost, a novel 3D-aware framework that is built upon diffusion models to perform instruction-based image composition. Previous methods concentrate on image compositing at the 2D level, which fall short in handling complex spatial relationships ($\textit{e.g.}$, occlusion). Bifr\"ost addresses these issues by training MLLM as a 2.5D location predictor and integrating depth maps as an extra condition during the generation process to bridge the gap between 2D and 3D, which enhances spatial comprehension and supports sophisticated spatial interactions. Our method begins by fine-tuning MLLM with a custom counterfactual dataset to predict 2.5D object locations in complex backgrounds from language instructions. Then, the image-compositing model is uniquely designed to process multiple types of input features, enabling it to perform high-fidelity image compositions that consider occlusion, depth blur, and image harmonization. Extensive qualitative and quantitative evaluations demonstrate that Bifr\"ost significantly outperforms existing methods, providing a robust solution for generating realistically composed images in scenarios demanding intricate spatial understanding. This work not only pushes the boundaries of generative image compositing but also reduces reliance on expensive annotated datasets by effectively utilizing existing resources in innovative ways.

VastTrack: Vast Category Visual Object Tracking

In this paper, we introduce a novel benchmark, dubbed VastTrack, towards facilitating the development of more general visual tracking via encompassing abundant classes and videos. VastTrack possesses several attractive properties: (1) Vast Object Category. In particular, it covers target objects from 2,115 classes, largely surpassing object categories of existing popular benchmarks (e.g., GOT-10k with 563 classes and LaSOT with 70 categories). With such vast object classes, we expect to learn more general object tracking. (2) Larger scale. Compared with current benchmarks, VastTrack offers 50,610 sequences with 4.2 million frames, which makes it to date the largest benchmark regarding the number of videos, and thus could benefit training even more powerful visual trackers in the deep learning era. (3) Rich Annotation. Besides conventional bounding box annotations, VastTrack also provides linguistic descriptions for the videos. The rich annotations of VastTrack enables development of both the vision-only and the vision-language tracking. To ensure precise annotation, all videos are manually labeled with multiple rounds of careful inspection and refinement. To understand performance of existing trackers and to provide baselines for future comparison, we extensively assess 25 representative trackers. The results, not surprisingly, show significant drops compared to those on current datasets due to lack of abundant categories and videos from diverse scenarios for training, and more efforts are required to improve general tracking. Our VastTrack and all the evaluation results will be made publicly available https://github.com/HengLan/VastTrack.

Deep Residual Networks with a Fully Connected Recon-struction Layer for Single Image Super-Resolution

Recently, deep neural networks have achieved impressive performance in terms of both reconstruction accuracy and efficiency for single image super-resolution (SISR). However, the network model of these methods is a fully convolutional neural network, which is limit to exploit contextual information over the global region of the input image. In this paper, we discuss a new SR architecture where features are extracted in the low-resolution (LR) space, and then we use a fully connected layer which learns an array of upsampling weights to reconstruct the desired high-resolution (HR) image from the final LR features. By doing so, we effectively exploit global context information over the input image region, whilst maintaining the low computational complexity for the overall SR operation. In addition, we introduce an edge difference constraint into our loss function to pre-serve edges and texture structures. Extensive experiments validate that our meth-od outperforms the existing state-of-the-art methods

Deep Inception-Residual Laplacian Pyramid Networks for Accurate Single Image Super-Resolution

With exploiting contextual information over large image regions in an efficient way, the deep convolutional neural network has shown an impressive performance for single image super-resolution (SR). In this paper, we propose a deep convolutional network by cascading the well-designed inception-residual blocks within the deep Laplacian pyramid framework to progressively restore the missing high-frequency details of high-resolution (HR) images. By optimizing our network structure, the trainable depth of the proposed network gains a significant improvement, which in turn improves super-resolving accuracy. With our network depth increasing, however, the saturation and degradation of training accuracy continues to be a critical problem. As regard to this, we propose an effective two-stage training strategy, in which we firstly use images downsampled from the ground-truth HR images as the optimal objective to train the inception-residual blocks in each pyramid level with an extremely high learning rate enabled by gradient clipping, and then the ground-truth HR images are used to fine-tune all the pre-trained inception-residual blocks for obtaining the final SR model. Furthermore, we present a new loss function operating in both image space and local rank space to optimize our network for exploiting the contextual information among different output components. Extensive experiments on benchmark datasets validate that the proposed method outperforms existing state-of-the-art SR methods in terms of the objective evaluation as well as the visual quality.