Predicting Long-horizon Futures by Conditioning on Geometry and Time

Our work explores the task of generating future sensor observations conditioned on the past. We are motivated by `predictive coding' concepts from neuroscience as well as robotic applications such as self-driving vehicles. Predictive video modeling is challenging because the future may be multi-modal and learning at scale remains computationally expensive for video processing. To address both challenges, our key insight is to leverage the large-scale pretraining of image diffusion models which can handle multi-modality. We repurpose image models for video prediction by conditioning on new frame timestamps. Such models can be trained with videos of both static and dynamic scenes. To allow them to be trained with modestly-sized datasets, we introduce invariances by factoring out illumination and texture by forcing the model to predict (pseudo) depth, readily obtained for in-the-wild videos via off-the-shelf monocular depth networks. In fact, we show that simply modifying networks to predict grayscale pixels already improves the accuracy of video prediction. Given the extra controllability with timestamp conditioning, we propose sampling schedules that work better than the traditional autoregressive and hierarchical sampling strategies. Motivated by probabilistic metrics from the object forecasting literature, we create a benchmark for video prediction on a diverse set of videos spanning indoor and outdoor scenes and a large vocabulary of objects. Our experiments illustrate the effectiveness of learning to condition on timestamps, and show the importance of predicting the future with invariant modalities.

Evaluating Text-to-Visual Generation with Image-to-Text Generation

Despite significant progress in generative AI, comprehensive evaluation remains challenging because of the lack of effective metrics and standardized benchmarks. For instance, the widely-used CLIPScore measures the alignment between a (generated) image and text prompt, but it fails to produce reliable scores for complex prompts involving compositions of objects, attributes, and relations. One reason is that text encoders of CLIP can notoriously act as a "bag of words", conflating prompts such as "the horse is eating the grass" with "the grass is eating the horse". To address this, we introduce the VQAScore, which uses a visual-question-answering (VQA) model to produce an alignment score by computing the probability of a "Yes" answer to a simple "Does this figure show '{text}'?" question. Though simpler than prior art, VQAScore computed with off-the-shelf models produces state-of-the-art results across many (8) image-text alignment benchmarks. We also compute VQAScore with an in-house model that follows best practices in the literature. For example, we use a bidirectional image-question encoder that allows image embeddings to depend on the question being asked (and vice versa). Our in-house model, CLIP-FlanT5, outperforms even the strongest baselines that make use of the proprietary GPT-4V. Interestingly, although we train with only images, VQAScore can also align text with video and 3D models. VQAScore allows researchers to benchmark text-to-visual generation using complex texts that capture the compositional structure of real-world prompts. We introduce GenAI-Bench, a more challenging benchmark with 1,600 compositional text prompts that require parsing scenes, objects, attributes, relationships, and high-order reasoning like comparison and logic. GenAI-Bench also offers over 15,000 human ratings for leading image and video generation models such as Stable Diffusion, DALL-E 3, and Gen2.

Better Call SAL: Towards Learning to Segment Anything in Lidar

We propose $\texttt{SAL}$ ($\texttt{S}$egment $\texttt{A}$nything in $\texttt{L}$idar) method consisting of a text-promptable zero-shot model for segmenting and classifying any object in Lidar, and a pseudo-labeling engine that facilitates model training without manual supervision. While the established paradigm for $\textit{Lidar Panoptic Segmentation}$ (LPS) relies on manual supervision for a handful of object classes defined a priori, we utilize 2D vision foundation models to generate 3D supervision "for free". Our pseudo-labels consist of instance masks and corresponding CLIP tokens, which we lift to Lidar using calibrated multi-modal data. By training our model on these labels, we distill the 2D foundation models into our Lidar $\texttt{SAL}$ model. Even without manual labels, our model achieves $91\%$ in terms of class-agnostic segmentation and $44\%$ in terms of zero-shot LPS of the fully supervised state-of-the-art. Furthermore, we outperform several baselines that do not distill but only lift image features to 3D. More importantly, we demonstrate that $\texttt{SAL}$ supports arbitrary class prompts, can be easily extended to new datasets, and shows significant potential to improve with increasing amounts of self-labeled data.

I Can't Believe It's Not Scene Flow!

Current scene flow methods broadly fail to describe motion on small objects, and current scene flow evaluation protocols hide this failure by averaging over many points, with most drawn larger objects. To fix this evaluation failure, we propose a new evaluation protocol, Bucket Normalized EPE, which is class-aware and speed-normalized, enabling contextualized error comparisons between object types that move at vastly different speeds. To highlight current method failures, we propose a frustratingly simple supervised scene flow baseline, TrackFlow, built by bolting a high-quality pretrained detector (trained using many class rebalancing techniques) onto a simple tracker, that produces state-of-the-art performance on current standard evaluations and large improvements over prior art on our new evaluation. Our results make it clear that all scene flow evaluations must be class and speed aware, and supervised scene flow methods must address point class imbalances. We release the evaluation code publicly at https://github.com/kylevedder/BucketedSceneFlowEval.

Cameras as Rays: Pose Estimation via Ray Diffusion

Estimating camera poses is a fundamental task for 3D reconstruction and remains challenging given sparse views (<10). In contrast to existing approaches that pursue top-down prediction of global parametrizations of camera extrinsics, we propose a distributed representation of camera pose that treats a camera as a bundle of rays. This representation allows for a tight coupling with spatial image features improving pose precision. We observe that this representation is naturally suited for set-level level transformers and develop a regression-based approach that maps image patches to corresponding rays. To capture the inherent uncertainties in sparse-view pose inference, we adapt this approach to learn a denoising diffusion model which allows us to sample plausible modes while improving performance. Our proposed methods, both regression- and diffusion-based, demonstrate state-of-the-art performance on camera pose estimation on CO3D while generalizing to unseen object categories and in-the-wild captures.

FlashTex: Fast Relightable Mesh Texturing with LightControlNet

Manually creating textures for 3D meshes is time-consuming, even for expert visual content creators. We propose a fast approach for automatically texturing an input 3D mesh based on a user-provided text prompt. Importantly, our approach disentangles lighting from surface material/reflectance in the resulting texture so that the mesh can be properly relit and rendered in any lighting environment. We introduce LightControlNet, a new text-to-image model based on the ControlNet architecture, which allows the specification of the desired lighting as a conditioning image to the model. Our text-to-texture pipeline then constructs the texture in two stages. The first stage produces a sparse set of visually consistent reference views of the mesh using LightControlNet. The second stage applies a texture optimization based on Score Distillation Sampling (SDS) that works with LightControlNet to increase the texture quality while disentangling surface material from lighting. Our pipeline is significantly faster than previous text-to-texture methods, while producing high-quality and relightable textures.

The Neglected Tails of Vision-Language Models

Vision-language models (VLMs) excel in zero-shot recognition but their performance varies greatly across different visual concepts. For example, although CLIP achieves impressive accuracy on ImageNet (60-80%), its performance drops below 10% for more than ten concepts like night snake, presumably due to their limited presence in the pretraining data. However, measuring the frequency of concepts in VLMs' large-scale datasets is challenging. We address this by using large language models (LLMs) to count the number of pretraining texts that contain synonyms of these concepts. Our analysis confirms that popular datasets, such as LAION, exhibit a long-tailed concept distribution, yielding biased performance in VLMs. We also find that downstream applications of VLMs, including visual chatbots (e.g., GPT-4V) and text-to-image models (e.g., Stable Diffusion), often fail to recognize or generate images of rare concepts identified by our method. To mitigate the imbalanced performance of zero-shot VLMs, we propose REtrieval-Augmented Learning (REAL). First, instead of prompting VLMs using the original class names, REAL uses their most frequent synonyms found in pretraining texts. This simple change already outperforms costly human-engineered and LLM-enriched prompts over nine benchmark datasets. Second, REAL trains a linear classifier on a small yet balanced set of pretraining data retrieved using concept synonyms. REAL surpasses the previous zero-shot SOTA, using 400x less storage and 10,000x less training time!

Revisiting Few-Shot Object Detection with Vision-Language Models

Few-shot object detection (FSOD) benchmarks have advanced techniques for detecting new categories with limited annotations. Existing benchmarks repurpose well-established datasets like COCO by partitioning categories into base and novel classes for pre-training and fine-tuning respectively. However, these benchmarks do not reflect how FSOD is deployed in practice. Rather than only pre-training on a small number of base categories, we argue that it is more practical to fine-tune a foundation model (e.g., a vision-language model (VLM) pre-trained on web-scale data) for a target domain. Surprisingly, we find that zero-shot inference from VLMs like GroundingDINO significantly outperforms the state-of-the-art (48.3 vs. 33.1 AP) on COCO. However, such zero-shot models can still be misaligned to target concepts of interest. For example, trailers on the web may be different from trailers in the context of autonomous vehicles. In this work, we propose Foundational FSOD, a new benchmark protocol that evaluates detectors pre-trained on any external datasets and fine-tuned on K-shots per target class. Further, we note that current FSOD benchmarks are actually federated datasets containing exhaustive annotations for each category on a subset of the data. We leverage this insight to propose simple strategies for fine-tuning VLMs with federated losses. We demonstrate the effectiveness of our approach on LVIS and nuImages, improving over prior work by 5.9 AP.

Tracking Any Object Amodally

Amodal perception, the ability to comprehend complete object structures from partial visibility, is a fundamental skill, even for infants. Its significance extends to applications like autonomous driving, where a clear understanding of heavily occluded objects is essential. However, modern detection and tracking algorithms often overlook this critical capability, perhaps due to the prevalence of modal annotations in most datasets. To address the scarcity of amodal data, we introduce the TAO-Amodal benchmark, featuring 880 diverse categories in thousands of video sequences. Our dataset includes amodal and modal bounding boxes for visible and occluded objects, including objects that are partially out-of-frame. To enhance amodal tracking with object permanence, we leverage a lightweight plug-in module, the amodal expander, to transform standard, modal trackers into amodal ones through fine-tuning on a few hundred video sequences with data augmentation. We achieve a 3.3\% and 1.6\% improvement on the detection and tracking of occluded objects on TAO-Amodal. When evaluated on people, our method produces dramatic improvements of 2x compared to state-of-the-art modal baselines.