Improving the detection of distant 3d objects is an important yet challenging task. For camera-based 3D perception, the annotation of 3d bounding relies heavily on LiDAR for accurate depth information. As such, the distance of annotation is often limited due to the sparsity of LiDAR points on distant objects, which hampers the capability of existing detectors for long-range scenarios. We address this challenge by considering only 2D box supervision for distant objects since they are easy to annotate. We propose LR3D, a framework that learns to recover the missing depth of distant objects. LR3D adopts an implicit projection head to learn the generation of mapping between 2D boxes and depth using the 3D supervision on close objects. This mapping allows the depth estimation of distant objects conditioned on their 2D boxes, making long-range 3D detection with 2D supervision feasible. Experiments show that without distant 3D annotations, LR3D allows camera-based methods to detect distant objects (over 200m) with comparable accuracy to full 3D supervision. Our framework is general, and could widely benefit 3D detection methods to a large extent.
In this work, we re-examine inter-patch dependencies in the decoding mechanism of masked autoencoders (MAE). We decompose this decoding mechanism for masked patch reconstruction in MAE into self-attention and cross-attention. Our investigations suggest that self-attention between mask patches is not essential for learning good representations. To this end, we propose a novel pretraining framework: Cross-Attention Masked Autoencoders (CrossMAE). CrossMAE's decoder leverages only cross-attention between masked and visible tokens, with no degradation in downstream performance. This design also enables decoding only a small subset of mask tokens, boosting efficiency. Furthermore, each decoder block can now leverage different encoder features, resulting in improved representation learning. CrossMAE matches MAE in performance with 2.5 to 3.7$\times$ less decoding compute. It also surpasses MAE on ImageNet classification and COCO instance segmentation under the same compute. Code and models: https://crossmae.github.io
Prior work has established test-time training (TTT) as a general framework to further improve a trained model at test time. Before making a prediction on each test instance, the model is trained on the same instance using a self-supervised task, such as image reconstruction with masked autoencoders. We extend TTT to the streaming setting, where multiple test instances - video frames in our case - arrive in temporal order. Our extension is online TTT: The current model is initialized from the previous model, then trained on the current frame and a small window of frames immediately before. Online TTT significantly outperforms the fixed-model baseline for four tasks, on three real-world datasets. The relative improvement is 45% and 66% for instance and panoptic segmentation. Surprisingly, online TTT also outperforms its offline variant that accesses more information, training on all frames from the entire test video regardless of temporal order. This differs from previous findings using synthetic videos. We conceptualize locality as the advantage of online over offline TTT. We analyze the role of locality with ablations and a theory based on bias-variance trade-off.
Robots operating in the real world require both rich manipulation skills as well as the ability to semantically reason about when to apply those skills. Towards this goal, recent works have integrated semantic representations from large-scale pretrained vision-language (VL) models into manipulation models, imparting them with more general reasoning capabilities. However, we show that the conventional pretraining-finetuning pipeline for integrating such representations entangles the learning of domain-specific action information and domain-general visual information, leading to less data-efficient training and poor generalization to unseen objects and tasks. To this end, we propose ProgramPort, a modular approach to better leverage pretrained VL models by exploiting the syntactic and semantic structures of language instructions. Our framework uses a semantic parser to recover an executable program, composed of functional modules grounded on vision and action across different modalities. Each functional module is realized as a combination of deterministic computation and learnable neural networks. Program execution produces parameters to general manipulation primitives for a robotic end-effector. The entire modular network can be trained with end-to-end imitation learning objectives. Experiments show that our model successfully disentangles action and perception, translating to improved zero-shot and compositional generalization in a variety of manipulation behaviors. Project webpage at: \url{https://progport.github.io}.
In recent years, increasing attention has been directed to leveraging pre-trained vision models for motor control. While existing works mainly emphasize the importance of this pre-training phase, the arguably equally important role played by downstream policy learning during control-specific fine-tuning is often neglected. It thus remains unclear if pre-trained vision models are consistent in their effectiveness under different control policies. To bridge this gap in understanding, we conduct a comprehensive study on 14 pre-trained vision models using 3 distinct classes of policy learning methods, including reinforcement learning (RL), imitation learning through behavior cloning (BC), and imitation learning with a visual reward function (VRF). Our study yields a series of intriguing results, including the discovery that the effectiveness of pre-training is highly dependent on the choice of the downstream policy learning algorithm. We show that conventionally accepted evaluation based on RL methods is highly variable and therefore unreliable, and further advocate for using more robust methods like VRF and BC. To facilitate more universal evaluations of pre-trained models and their policy learning methods in the future, we also release a benchmark of 21 tasks across 3 different environments alongside our work.
Humans are remarkably good at understanding and reasoning about complex visual scenes. The capability to decompose low-level observations into discrete objects allows us to build a grounded abstract representation and identify the compositional structure of the world. Accordingly, it is a crucial step for machine learning models to be capable of inferring objects and their properties from visual scenes without explicit supervision. However, existing works on object-centric representation learning either rely on tailor-made neural network modules or strong probabilistic assumptions in the underlying generative and inference processes. In this work, we present \ours, a conceptually simple and general approach to learning object-centric representations through an energy-based model. By forming a permutation-invariant energy function using vanilla attention blocks readily available in Transformers, we can infer object-centric latent variables via gradient-based MCMC methods where permutation equivariance is automatically guaranteed. We show that \ours can be easily integrated into existing architectures and can effectively extract high-quality object-centric representations, leading to better segmentation accuracy and competitive downstream task performance. Further, empirical evaluations show that \ours's learned representations are robust against distribution shift. Finally, we demonstrate the effectiveness of \ours in systematic compositional generalization, by re-composing learned energy functions for novel scene generation and manipulation.
Establishing visual correspondence across images is a challenging and essential task. Recently, an influx of self-supervised methods have been proposed to better learn representations for visual correspondence. However, we find that these methods often fail to leverage semantic information and over-rely on the matching of low-level features. In contrast, human vision is capable of distinguishing between distinct objects as a pretext to tracking. Inspired by this paradigm, we propose to learn semantic-aware fine-grained correspondence. Firstly, we demonstrate that semantic correspondence is implicitly available through a rich set of image-level self-supervised methods. We further design a pixel-level self-supervised learning objective which specifically targets fine-grained correspondence. For downstream tasks, we fuse these two kinds of complementary correspondence representations together, demonstrating that they boost performance synergistically. Our method surpasses previous state-of-the-art self-supervised methods using convolutional networks on a variety of visual correspondence tasks, including video object segmentation, human pose tracking, and human part tracking.
Many recent approaches in contrastive learning have worked to close the gap between pretraining on iconic images like ImageNet and pretraining on complex scenes like COCO. This gap exists largely because commonly used random crop augmentations obtain semantically inconsistent content in crowded scene images of diverse objects. Previous works use preprocessing pipelines to localize salient objects for improved cropping, but an end-to-end solution is still elusive. In this work, we propose a framework which accomplishes this goal via joint learning of representations and segmentation. We leverage segmentation masks to train a model with a mask-dependent contrastive loss, and use the partially trained model to bootstrap better masks. By iterating between these two components, we ground the contrastive updates in segmentation information, and simultaneously improve segmentation throughout pretraining. Experiments show our representations transfer robustly to downstream tasks in classification, detection and segmentation.
Accurate forecasts of fine particulate matter (PM 2.5) from wildfire smoke are crucial to safeguarding cardiopulmonary public health. Existing forecasting systems are trained on sparse and inaccurate ground truths, and do not take sufficient advantage of important spatial inductive biases. In this work, we present a convolutional neural network which preserves sparsity invariance throughout, and leverages multitask learning to perform dense forecasts of PM 2.5values. We demonstrate that our model outperforms two existing smoke forecasting systems during the 2018 and 2019 wildfire season in British Columbia, Canada, predicting PM 2.5 at a grid resolution of 10 km, 24 hours in advance with high fidelity. Most interestingly, our model also generalizes to meaningful smoke dispersion patterns despite training with irregularly distributed ground truth PM 2.5 values available in only 0.5% of grid cells.
In many real-world settings, we are interested in learning invariant and equivariant functions over nested or multiresolution structures, such as a set of sequences, a graph of graphs, or a multiresolution image. While equivariant linear maps and by extension multilayer perceptrons (MLPs) for many of the individual basic structures are known, a formalism for dealing with a hierarchy of symmetry transformations is lacking. Observing that the transformation group for a nested structure corresponds to the ``wreath product'' of the symmetry groups of the building blocks, we show how to obtain the equivariant map for hierarchical data-structures using an intuitive combination of the equivariant maps for the individual blocks. To demonstrate the effectiveness of this type of model, we use a hierarchy of translation and permutation symmetries for learning on point cloud data, and report state-of-the-art on \kw{semantic3d} and \kw{s3dis}, two of the largest real-world benchmarks for 3D semantic segmentation.