Zaiwei Zhang

Zhenpei Yang, Zhile Ren, Miguel Angel Bautista et al. · apple-ml

Reconstructing an accurate 3D object model from a few image observations remains a challenging problem in computer vision. State-of-the-art approaches typically assume accurate camera poses as input, which could be difficult to obtain in realistic settings. In this paper, we present FvOR, a learning-based object reconstruction method that predicts accurate 3D models given a few images with noisy input poses. The core of our approach is a fast and robust multi-view reconstruction algorithm to jointly refine 3D geometry and camera pose estimation using learnable neural network modules. We provide a thorough benchmark of state-of-the-art approaches for this problem on ShapeNet. Our approach achieves best-in-class results. It is also two orders of magnitude faster than the recent optimization-based approach IDR. Our code is released at \url{https://github.com/zhenpeiyang/FvOR/}

Zhenpei Yang, Zaiwei Zhang, Qixing Huang

Reconstructing 3D objects is an important computer vision task that has wide application in AR/VR. Deep learning algorithm developed for this task usually relies on an unrealistic synthetic dataset, such as ShapeNet and Things3D. On the other hand, existing real-captured object-centric datasets usually do not have enough annotation to enable supervised training or reliable evaluation. In this technical report, we present a photo-realistic object-centric dataset HM3D-ABO. It is constructed by composing realistic indoor scene and realistic object. For each configuration, we provide multi-view RGB observations, a water-tight mesh model for the object, ground truth depth map and object mask. The proposed dataset could also be useful for tasks such as camera pose estimation and novel-view synthesis. The dataset generation code is released at https://github.com/zhenpeiyang/HM3D-ABO.

Chenwei Wu, Li Erran Li, Stefano Ermon et al.

Compositionality is a common property in many modalities including natural languages and images, but the compositional generalization of multi-modal models is not well-understood. In this paper, we identify two sources of visual-linguistic compositionality: linguistic priors and the interplay between images and texts. We show that current attempts to improve compositional generalization rely on linguistic priors rather than on information in the image. We also propose a new metric for compositionality without such linguistic priors.

Haitao Yang, Zaiwei Zhang, Xiangru Huang et al.

Bird's-Eye View (BEV) features are popular intermediate scene representations shared by the 3D backbone and the detector head in LiDAR-based object detectors. However, little research has been done to investigate how to incorporate additional supervision on the BEV features to improve proposal generation in the detector head, while still balancing the number of powerful 3D layers and efficient 2D network operations. This paper proposes a novel scene representation that encodes both the semantics and geometry of the 3D environment in 2D, which serves as a dense supervision signal for better BEV feature learning. The key idea is to use auxiliary networks to predict a combination of explicit and implicit semantic probabilities by exploiting their complementary properties. Extensive experiments show that our simple yet effective design can be easily integrated into most state-of-the-art 3D object detectors and consistently improves upon baseline models.

Mao Ye, Gregory P. Meyer, Zaiwei Zhang et al.

Ensuring robust performance on long-tail examples is an important problem for many real-world applications of machine learning, such as autonomous driving. This work focuses on the problem of identifying rare examples within a corpus of unlabeled data. We propose a simple and scalable data mining approach that leverages the knowledge contained within a large vision language model (VLM). Our approach utilizes a VLM to summarize the content of an image into a set of keywords, and we identify rare examples based on keyword frequency. We find that the VLM offers a distinct signal for identifying long-tail examples when compared to conventional methods based on model uncertainty. Therefore, we propose a simple and general approach for integrating signals from multiple mining algorithms. We evaluate the proposed method on two diverse tasks: 2D image classification, in which inter-class variation is the primary source of data diversity, and on 3D object detection, where intra-class variation is the main concern. Furthermore, through the detection task, we demonstrate that the knowledge extracted from 2D images is transferable to the 3D domain. Our experiments consistently show large improvements (between 10\% and 50\%) over the baseline techniques on several representative benchmarks: ImageNet-LT, Places-LT, and the Waymo Open Dataset.

Zaiwei Zhang, Gregory P. Meyer, Zhichao Lu et al.

For visual recognition, knowledge distillation typically involves transferring knowledge from a large, well-trained teacher model to a smaller student model. In this paper, we introduce an effective method to distill knowledge from an off-the-shelf vision-language model (VLM), demonstrating that it provides novel supervision in addition to those from a conventional vision-only teacher model. Our key technical contribution is the development of a framework that generates novel text supervision and distills free-form text into a vision encoder. We showcase the effectiveness of our approach, termed VLM-KD, across various benchmark datasets, showing that it surpasses several state-of-the-art long-tail visual classifiers. To our knowledge, this work is the first to utilize knowledge distillation with text supervision generated by an off-the-shelf VLM and apply it to vanilla randomly initialized vision encoders.

Yi Xu, Yuxin Hu, Zaiwei Zhang et al.

Human drivers rely on commonsense reasoning to navigate diverse and dynamic real-world scenarios. Existing end-to-end (E2E) autonomous driving (AD) models are typically optimized to mimic driving patterns observed in data, without capturing the underlying reasoning processes. This limitation constrains their ability to handle challenging driving scenarios. To close this gap, we propose VLM-AD, a method that leverages vision-language models (VLMs) as teachers to enhance training by providing additional supervision that incorporates unstructured reasoning information and structured action labels. Such supervision enhances the model's ability to learn richer feature representations that capture the rationale behind driving patterns. Importantly, our method does not require a VLM during inference, making it practical for real-time deployment. When integrated with state-of-the-art methods, VLM-AD achieves significant improvements in planning accuracy and reduced collision rates on the nuScenes dataset. It further improves route completion and driving scores under closed-loop evaluation, demonstrating its effectiveness in long-horizon, interactive driving scenarios and its potential for safe and reliable real-world deployment.

Liyan Chen, Gregory P. Meyer, Zaiwei Zhang et al.

Recent efforts recognize the power of scale in 3D learning (e.g. PTv3) and attention mechanisms (e.g. FlashAttention). However, current point cloud backbones fail to holistically unify geometric locality, attention mechanisms, and GPU architectures in one view. In this paper, we introduce Flash3D Transformer, which aligns geometric locality and GPU tiling through a principled locality mechanism based on Perfect Spatial Hashing (PSH). The common alignment with GPU tiling naturally fuses our PSH locality mechanism with FlashAttention at negligible extra cost. This mechanism affords flexible design choices throughout the backbone that result in superior downstream task results. Flash3D outperforms state-of-the-art PTv3 results on benchmark datasets, delivering a 2.25x speed increase and 2.4x memory efficiency boost. This efficiency enables scaling to wider attention scopes and larger models without additional overhead. Such scaling allows Flash3D to achieve even higher task accuracies than PTv3 under the same compute budget.

Haitao Yang, Zaiwei Zhang, Siming Yan et al.

Developing deep neural networks to generate 3D scenes is a fundamental problem in neural synthesis with immediate applications in architectural CAD, computer graphics, as well as in generating virtual robot training environments. This task is challenging because 3D scenes exhibit diverse patterns, ranging from continuous ones, such as object sizes and the relative poses between pairs of shapes, to discrete patterns, such as occurrence and co-occurrence of objects with symmetrical relationships. This paper introduces a novel neural scene synthesis approach that can capture diverse feature patterns of 3D scenes. Our method combines the strength of both neural network-based and conventional scene synthesis approaches. We use the parametric prior distributions learned from training data, which provide uncertainties of object attributes and relative attributes, to regularize the outputs of feed-forward neural models. Moreover, instead of merely predicting a scene layout, our approach predicts an over-complete set of attributes. This methodology allows us to utilize the underlying consistency constraints among the predicted attributes to prune infeasible predictions. Experimental results show that our approach outperforms existing methods considerably. The generated 3D scenes interpolate the training data faithfully while preserving both continuous and discrete feature patterns.

Qixing Huang, Xiangru Huang, Bo Sun et al.

This paper introduces an unsupervised loss for training parametric deformation shape generators. The key idea is to enforce the preservation of local rigidity among the generated shapes. Our approach builds on an approximation of the as-rigid-as possible (or ARAP) deformation energy. We show how to develop the unsupervised loss via a spectral decomposition of the Hessian of the ARAP energy. Our loss nicely decouples pose and shape variations through a robust norm. The loss admits simple closed-form expressions. It is easy to train and can be plugged into any standard generation models, e.g., variational auto-encoder (VAE) and auto-decoder (AD). Experimental results show that our approach outperforms existing shape generation approaches considerably on public benchmark datasets of various shape categories such as human, animal and bone.

Zaiwei Zhang, Rohit Girdhar, Armand Joulin et al.

Pretraining on large labeled datasets is a prerequisite to achieve good performance in many computer vision tasks like 2D object recognition, video classification etc. However, pretraining is not widely used for 3D recognition tasks where state-of-the-art methods train models from scratch. A primary reason is the lack of large annotated datasets because 3D data is both difficult to acquire and time consuming to label. We present a simple self-supervised pertaining method that can work with any 3D data - single or multiview, indoor or outdoor, acquired by varied sensors, without 3D registration. We pretrain standard point cloud and voxel based model architectures, and show that joint pretraining further improves performance. We evaluate our models on 9 benchmarks for object detection, semantic segmentation, and object classification, where they achieve state-of-the-art results and can outperform supervised pretraining. We set a new state-of-the-art for object detection on ScanNet (69.0% mAP) and SUNRGBD (63.5% mAP). Our pretrained models are label efficient and improve performance for classes with few examples.

Zaiwei Zhang, Bo Sun, Haitao Yang et al.

We introduce H3DNet, which takes a colorless 3D point cloud as input and outputs a collection of oriented object bounding boxes (or BB) and their semantic labels. The critical idea of H3DNet is to predict a hybrid set of geometric primitives, i.e., BB centers, BB face centers, and BB edge centers. We show how to convert the predicted geometric primitives into object proposals by defining a distance function between an object and the geometric primitives. This distance function enables continuous optimization of object proposals, and its local minimums provide high-fidelity object proposals. H3DNet then utilizes a matching and refinement module to classify object proposals into detected objects and fine-tune the geometric parameters of the detected objects. The hybrid set of geometric primitives not only provides more accurate signals for object detection than using a single type of geometric primitives, but it also provides an overcomplete set of constraints on the resulting 3D layout. Therefore, H3DNet can tolerate outliers in predicted geometric primitives. Our model achieves state-of-the-art 3D detection results on two large datasets with real 3D scans, ScanNet and SUN RGB-D.

Zaiwei Zhang, Xiangru Huang, Qixing Huang et al.

In this paper, we introduce the problem of jointly learning feed-forward neural networks across a set of relevant but diverse datasets. Compared to learning a separate network from each dataset in isolation, joint learning enables us to extract correlated information across multiple datasets to significantly improve the quality of learned networks. We formulate this problem as joint learning of multiple copies of the same network architecture and enforce the network weights to be shared across these networks. Instead of hand-encoding the shared network layers, we solve an optimization problem to automatically determine how layers should be shared between each pair of datasets. Experimental results show that our approach outperforms baselines without joint learning and those using pretraining-and-fine-tuning. We show the effectiveness of our approach on three tasks: image classification, learning auto-encoders, and image generation.

Zaiwei Zhang, Zhenxiao Liang, Lemeng Wu et al.

Optimizing a network of maps among a collection of objects/domains (or map synchronization) is a central problem across computer vision and many other relevant fields. Compared to optimizing pairwise maps in isolation, the benefit of map synchronization is that there are natural constraints among a map network that can improve the quality of individual maps. While such self-supervision constraints are well-understood for undirected map networks (e.g., the cycle-consistency constraint), they are under-explored for directed map networks, which naturally arise when maps are given by parametric maps (e.g., a feed-forward neural network). In this paper, we study a natural self-supervision constraint for directed map networks called path-invariance, which enforces that composite maps along different paths between a fixed pair of source and target domains are identical. We introduce path-invariance bases for efficient encoding of the path-invariance constraint and present an algorithm that outputs a path-variance basis with polynomial time and space complexities. We demonstrate the effectiveness of our approach on optimizing object correspondences, estimating dense image maps via neural networks, and semantic segmentation of 3D scenes via map networks of diverse 3D representations. In particular, for 3D semantic segmentation, our approach only requires 8% labeled data from ScanNet to achieve the same performance as training a single 3D segmentation network with 30% to 100% labeled data.

Zaiwei Zhang, Zhenpei Yang, Chongyang Ma et al.

We present a deep generative scene modeling technique for indoor environments. Our goal is to train a generative model using a feed-forward neural network that maps a prior distribution (e.g., a normal distribution) to the distribution of primary objects in indoor scenes. We introduce a 3D object arrangement representation that models the locations and orientations of objects, based on their size and shape attributes. Moreover, our scene representation is applicable for 3D objects with different multiplicities (repetition counts), selected from a database. We show a principled way to train this model by combining discriminator losses for both a 3D object arrangement representation and a 2D image-based representation. We demonstrate the effectiveness of our scene representation and the deep learning method on benchmark datasets. We also show the applications of this generative model in scene interpolation and scene completion.