The scaling of Transformers has driven breakthrough capabilities for language models. At present, the largest large language models (LLMs) contain upwards of 100B parameters. Vision Transformers (ViT) have introduced the same architecture to image and video modelling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters (Chen et al., 2022). We present a recipe for highly efficient and stable training of a 22B-parameter ViT (ViT-22B) and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features), ViT-22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT-22B demonstrates the potential for "LLM-like" scaling in vision, and provides key steps towards getting there.
This work introduces an effective and practical solution to the dense two-view structure from motion (SfM) problem. One vital question addressed is how to mindfully use per-pixel optical flow correspondence between two frames for accurate pose estimation -- as perfect per-pixel correspondence between two images is difficult, if not impossible, to establish. With the carefully estimated camera pose and predicted per-pixel optical flow correspondences, a dense depth of the scene is computed. Later, an iterative refinement procedure is introduced to further improve optical flow matching confidence, camera pose, and depth, exploiting their inherent dependency in rigid SfM. The fundamental idea presented is to benefit from per-pixel uncertainty in the optical flow estimation and provide robustness to the dense SfM system via an online refinement. Concretely, we introduce a pipeline consisting of (i) an uncertainty-aware dense optical flow estimation approach that provides per-pixel correspondence with their confidence score of matching; (ii) a weighted dense bundle adjustment formulation that depends on optical flow uncertainty and bidirectional optical flow consistency to refine both pose and depth; (iii) a depth estimation network that considers its consistency with the estimated poses and optical flow respecting epipolar constraint. Extensive experiments show that the proposed approach achieves remarkable depth accuracy and state-of-the-art camera pose results superseding SuperPoint and SuperGlue accuracy when tested on benchmark datasets such as DeMoN, YFCC100M, and ScanNet.
Network binarization emerges as one of the most promising compression approaches offering extraordinary computation and memory savings by minimizing the bit-width. However, recent research has shown that applying existing binarization algorithms to diverse tasks, architectures, and hardware in realistic scenarios is still not straightforward. Common challenges of binarization, such as accuracy degradation and efficiency limitation, suggest that its attributes are not fully understood. To close this gap, we present BiBench, a rigorously designed benchmark with in-depth analysis for network binarization. We first carefully scrutinize the requirements of binarization in the actual production and define evaluation tracks and metrics for a comprehensive and fair investigation. Then, we evaluate and analyze a series of milestone binarization algorithms that function at the operator level and with extensive influence. Our benchmark reveals that 1) the binarized operator has a crucial impact on the performance and deployability of binarized networks; 2) the accuracy of binarization varies significantly across different learning tasks and neural architectures; 3) binarization has demonstrated promising efficiency potential on edge devices despite the limited hardware support. The results and analysis also lead to a promising paradigm for accurate and efficient binarization. We believe that BiBench will contribute to the broader adoption of binarization and serve as a foundation for future research.
To track the 3D locations and trajectories of the other traffic participants at any given time, modern autonomous vehicles are equipped with multiple cameras that cover the vehicle's full surroundings. Yet, camera-based 3D object tracking methods prioritize optimizing the single-camera setup and resort to post-hoc fusion in a multi-camera setup. In this paper, we propose a method for panoramic 3D object tracking, called CC-3DT, that associates and models object trajectories both temporally and across views, and improves the overall tracking consistency. In particular, our method fuses 3D detections from multiple cameras before association, reducing identity switches significantly and improving motion modeling. Our experiments on large-scale driving datasets show that fusion before association leads to a large margin of improvement over post-hoc fusion. We set a new state-of-the-art with 12.6% improvement in average multi-object tracking accuracy (AMOTA) among all camera-based methods on the competitive NuScenes 3D tracking benchmark, outperforming previously published methods by 6.5% in AMOTA with the same 3D detector.
Multi-camera 3D object detection, a critical component for vision-only driving systems, has achieved impressive progress. Notably, transformer-based methods with 2D features augmented by 3D positional encodings (PE) have enjoyed great success. However, the mechanism and options of 3D PE have not been thoroughly explored. In this paper, we first explore, analyze and compare various 3D positional encodings. In particular, we devise 3D point PE and show its superior performance since more precise positioning may lead to superior 3D detection. In practice, we utilize monocular depth estimation to obtain the 3D point positions for multi-camera 3D object detection. The PE with estimated 3D point locations can bring significant improvements compared to the commonly used camera-ray PE. Among DETR-based strategies, our method achieves state-of-the-art 45.6 mAP and 55.1 NDS on the competitive nuScenes valuation set. It's the first time that the performance gap between the vision-only (DETR-based) and LiDAR-based methods is reduced within 5\% mAP and 6\% NDS.
We present a unified formulation and model for three motion and 3D perception tasks: optical flow, rectified stereo matching and unrectified stereo depth estimation from posed images. Unlike previous specialized architectures for each specific task, we formulate all three tasks as a unified dense correspondence matching problem, which can be solved with a single model by directly comparing feature similarities. Such a formulation calls for discriminative feature representations, which we achieve using a Transformer, in particular the cross-attention mechanism. We demonstrate that cross-attention enables integration of knowledge from another image via cross-view interactions, which greatly improves the quality of the extracted features. Our unified model naturally enables cross-task transfer since the model architecture and parameters are shared across tasks. We outperform RAFT with our unified model on the challenging Sintel dataset, and our final model that uses a few additional task-specific refinement steps outperforms or compares favorably to recent state-of-the-art methods on 10 popular flow, stereo and depth datasets, while being simpler and more efficient in terms of model design and inference speed.
Improving model's generalizability against domain shifts is crucial, especially for safety-critical applications such as autonomous driving. Real-world domain styles can vary substantially due to environment changes and sensor noises, but deep models only know the training domain style. Such domain style gap impedes model generalization on diverse real-world domains. Our proposed Normalization Perturbation (NP) can effectively overcome this domain style overfitting problem. We observe that this problem is mainly caused by the biased distribution of low-level features learned in shallow CNN layers. Thus, we propose to perturb the channel statistics of source domain features to synthesize various latent styles, so that the trained deep model can perceive diverse potential domains and generalizes well even without observations of target domain data in training. We further explore the style-sensitive channels for effective style synthesis. Normalization Perturbation only relies on a single source domain and is surprisingly effective and extremely easy to implement. Extensive experiments verify the effectiveness of our method for generalizing models under real-world domain shifts.
Autonomous drones can operate in remote and unstructured environments, enabling various real-world applications. However, the lack of effective vision-based algorithms has been a stumbling block to achieving this goal. Existing systems often require hand-engineered components for state estimation, planning, and control. Such a sequential design involves laborious tuning, human heuristics, and compounding delays and errors. This paper tackles the vision-based autonomous-drone-racing problem by learning deep sensorimotor policies. We use contrastive learning to extract robust feature representations from the input images and leverage a two-stage learning-by-cheating framework for training a neural network policy. The resulting policy directly infers control commands with feature representations learned from raw images, forgoing the need for globally-consistent state estimation, trajectory planning, and handcrafted control design. Our experimental results indicate that our vision-based policy can achieve the same level of racing performance as the state-based policy while being robust against different visual disturbances and distractors. We believe this work serves as a stepping-stone toward developing intelligent vision-based autonomous systems that control the drone purely from image inputs, like human pilots.
Multi-task learning promises better model generalization on a target task by jointly optimizing it with an auxiliary task. However, the current practice requires additional labeling efforts for the auxiliary task, while not guaranteeing better model performance. In this paper, we find that jointly training a dense prediction (target) task with a self-supervised (auxiliary) task can consistently improve the performance of the target task, while eliminating the need for labeling auxiliary tasks. We refer to this joint training as Composite Learning (CompL). Experiments of CompL on monocular depth estimation, semantic segmentation, and boundary detection show consistent performance improvements in fully and partially labeled datasets. Further analysis on depth estimation reveals that joint training with self-supervision outperforms most labeled auxiliary tasks. We also find that CompL can improve model robustness when the models are evaluated in new domains. These results demonstrate the benefits of self-supervision as an auxiliary task, and establish the design of novel task-specific self-supervised methods as a new axis of investigation for future multi-task learning research.
Similarity learning has been recognized as a crucial step for object tracking. However, existing multiple object tracking methods only use sparse ground truth matching as the training objective, while ignoring the majority of the informative regions in images. In this paper, we present Quasi-Dense Similarity Learning, which densely samples hundreds of object regions on a pair of images for contrastive learning. We combine this similarity learning with multiple existing object detectors to build Quasi-Dense Tracking (QDTrack), which does not require displacement regression or motion priors. We find that the resulting distinctive feature space admits a simple nearest neighbor search at inference time for object association. In addition, we show that our similarity learning scheme is not limited to video data, but can learn effective instance similarity even from static input, enabling a competitive tracking performance without training on videos or using tracking supervision. We conduct extensive experiments on a wide variety of popular MOT benchmarks. We find that, despite its simplicity, QDTrack rivals the performance of state-of-the-art tracking methods on all benchmarks and sets a new state-of-the-art on the large-scale BDD100K MOT benchmark, while introducing negligible computational overhead to the detector.