Vision Language Models (VLMs), which extend Large Language Models (LLM) by incorporating visual understanding capability, have demonstrated significant advancements in addressing open-ended visual question-answering (VQA) tasks. However, these models cannot accurately interpret images infused with text, a common occurrence in real-world scenarios. Standard procedures for extracting information from images often involve learning a fixed set of query embeddings. These embeddings are designed to encapsulate image contexts and are later used as soft prompt inputs in LLMs. Yet, this process is limited to the token count, potentially curtailing the recognition of scenes with text-rich context. To improve upon them, the present study introduces BLIVA: an augmented version of InstructBLIP with Visual Assistant. BLIVA incorporates the query embeddings from InstructBLIP and also directly projects encoded patch embeddings into the LLM, a technique inspired by LLaVA. This approach assists the model to capture intricate details potentially missed during the query decoding process. Empirical evidence demonstrates that our model, BLIVA, significantly enhances performance in processing text-rich VQA benchmarks (up to 17.76% in OCR-VQA benchmark) and in undertaking general (not particularly text-rich) VQA benchmarks (up to 7.9% in Visual Spatial Reasoning benchmark), and achieved 17.72% overall improvement in a comprehensive multimodal LLM benchmark (MME), comparing to our baseline InstructBLIP. BLIVA demonstrates significant capability in decoding real-world images, irrespective of text presence. To demonstrate the broad industry applications enabled by BLIVA, we evaluate the model using a new dataset comprising YouTube thumbnails paired with question-answer sets across 11 diverse categories. Our code and models are freely accessible at https://github.com/mlpc-ucsd/BLIVA.
3D scene understanding has gained significant attention due to its wide range of applications. However, existing methods for 3D scene understanding are limited to specific downstream tasks, which hinders their practicality in real-world applications. This paper presents Chat-3D, which combines the 3D visual perceptual ability of pre-trained 3D representations and the impressive reasoning and conversation capabilities of advanced LLMs to achieve the first universal dialogue systems for 3D scenes. Specifically, we align 3D representations into the feature space of LLMs, thus enabling LLMs to perceive the 3D world. Given the scarcity of 3D scene-text data, we propose a three-stage training strategy to efficiently utilize the available data for better alignment. To enhance the reasoning ability and develop a user-friendly interaction scheme, we further construct a high-quality object-centric 3D instruction dataset and design an associated object-centric prompt. Our experiments show that Chat-3D achieves an impressive ability to comprehend diverse instructions for 3D scenes, engage in intricate spatial reasoning, and incorporate external knowledge into its responses. Chat-3D achieves a 75.6% relative score compared with GPT-4 on the constructed instruction dataset.
While many natural language inference (NLI) datasets target certain semantic phenomena, e.g., negation, tense & aspect, monotonicity, and presupposition, to the best of our knowledge, there is no NLI dataset that involves diverse types of spatial expressions and reasoning. We fill this gap by semi-automatically creating an NLI dataset for spatial reasoning, called SpaceNLI. The data samples are automatically generated from a curated set of reasoning patterns, where the patterns are annotated with inference labels by experts. We test several SOTA NLI systems on SpaceNLI to gauge the complexity of the dataset and the system's capacity for spatial reasoning. Moreover, we introduce a Pattern Accuracy and argue that it is a more reliable and stricter measure than the accuracy for evaluating a system's performance on pattern-based generated data samples. Based on the evaluation results we find that the systems obtain moderate results on the spatial NLI problems but lack consistency per inference pattern. The results also reveal that non-projective spatial inferences (especially due to the "between" preposition) are the most challenging ones.
3D perceptual representations are well suited for robot manipulation as they easily encode occlusions and simplify spatial reasoning. Many manipulation tasks require high spatial precision in end-effector pose prediction, which typically demands high-resolution 3D feature grids that are computationally expensive to process. As a result, most manipulation policies operate directly in 2D, foregoing 3D inductive biases. In this paper, we introduce Act3D, a manipulation policy transformer that represents the robot's workspace using a 3D feature field with adaptive resolutions dependent on the task at hand. The model lifts 2D pre-trained features to 3D using sensed depth, and attends to them to compute features for sampled 3D points. It samples 3D point grids in a coarse to fine manner, featurizes them using relative-position attention, and selects where to focus the next round of point sampling. In this way, it efficiently computes 3D action maps of high spatial resolution. Act3D sets a new state-of-the-art in RL-Bench, an established manipulation benchmark, where it achieves 10% absolute improvement over the previous SOTA 2D multi-view policy on 74 RLBench tasks and 22% absolute improvement with 3x less compute over the previous SOTA 3D policy. We quantify the importance of relative spatial attention, large-scale vision-language pre-trained 2D backbones, and weight tying across coarse-to-fine attentions in ablative experiments. Code and videos are available on our project website: https://act3d.github.io/.
Robots rely heavily on sensors, especially RGB and depth cameras, to perceive and interact with the world. RGB cameras record 2D images with rich semantic information while missing precise spatial information. On the other side, depth cameras offer critical 3D geometry data but capture limited semantics. Therefore, integrating both modalities is crucial for learning representations for robotic perception and control. However, current research predominantly focuses on only one of these modalities, neglecting the benefits of incorporating both. To this end, we present $\textbf{Semantic-Geometric Representation} (\textbf{SGR})$, a universal perception module for robotics that leverages the rich semantic information of large-scale pre-trained 2D models and inherits the merits of 3D spatial reasoning. Our experiments demonstrate that SGR empowers the agent to successfully complete a diverse range of simulated and real-world robotic manipulation tasks, outperforming state-of-the-art methods significantly in both single-task and multi-task settings. Furthermore, SGR possesses the capability to generalize to novel semantic attributes, setting it apart from the other methods. Project website: https://semantic-geometric-representation.github.io.
Generative neural models hold great promise in enhancing programming education by synthesizing new content. We seek to design neural models that can automatically generate programming tasks for a given specification in the context of visual programming domains. Despite the recent successes of large generative models like GPT-4, our initial results show that these models are ineffective in synthesizing visual programming tasks and struggle with logical and spatial reasoning. We propose a novel neuro-symbolic technique, NeurTaskSyn, that can synthesize programming tasks for a specification given in the form of desired programming concepts exercised by its solution code and constraints on the visual task. NeurTaskSyn has two components: the first component is trained via imitation learning procedure to generate possible solution codes, and the second component is trained via reinforcement learning procedure to guide an underlying symbolic execution engine that generates visual tasks for these codes. We demonstrate the effectiveness of NeurTaskSyn through an extensive empirical evaluation and a qualitative study on reference tasks taken from the Hour of Code: Classic Maze challenge by Code-dot-org and the Intro to Programming with Karel course by CodeHS-dot-com.
Recent advancements in text-to-image diffusion models have yielded impressive results in generating realistic and diverse images. However, these models still struggle with complex prompts, such as those that involve numeracy and spatial reasoning. This work proposes to enhance prompt understanding capabilities in diffusion models. Our method leverages a pretrained large language model (LLM) for grounded generation in a novel two-stage process. In the first stage, the LLM generates a scene layout that comprises captioned bounding boxes from a given prompt describing the desired image. In the second stage, a novel controller guides an off-the-shelf diffusion model for layout-grounded image generation. Both stages utilize existing pretrained models without additional model parameter optimization. Our method significantly outperforms the base diffusion model and several strong baselines in accurately generating images according to prompts that require various capabilities, doubling the generation accuracy across four tasks on average. Furthermore, our method enables instruction-based multi-round scene specification and can handle prompts in languages not supported by the underlying diffusion model. We anticipate that our method will unleash users' creativity by accurately following more complex prompts. Our code, demo, and benchmark are available at: https://llm-grounded-diffusion.github.io
The recent GPT-4 has demonstrated extraordinary multi-modal abilities, such as directly generating websites from handwritten text and identifying humorous elements within images. These features are rarely observed in previous vision-language models. However, the technical details behind GPT-4 continue to remain undisclosed. We believe that the enhanced multi-modal generation capabilities of GPT-4 stem from the utilization of sophisticated large language models (LLM). To examine this phenomenon, we present MiniGPT-4, which aligns a frozen visual encoder with a frozen advanced LLM, Vicuna, using one projection layer. Our work, for the first time, uncovers that properly aligning the visual features with an advanced large language model can possess numerous advanced multi-modal abilities demonstrated by GPT-4, such as detailed image description generation and website creation from hand-drawn drafts. Furthermore, we also observe other emerging capabilities in MiniGPT-4, including writing stories and poems inspired by given images, teaching users how to cook based on food photos, and so on. In our experiment, we found that the model trained on short image caption pairs could produce unnatural language outputs (e.g., repetition and fragmentation). To address this problem, we curate a detailed image description dataset in the second stage to finetune the model, which consequently improves the model's generation reliability and overall usability. Our code, pre-trained model, and collected dataset are available at https://minigpt-4.github.io/.
The potential of large language models (LLMs) to reason like humans has been a highly contested topic in Machine Learning communities. However, the reasoning abilities of humans are multifaceted and can be seen in various forms, including analogical, spatial and moral reasoning, among others. This fact raises the question whether LLMs can perform equally well across all these different domains. This research work aims to investigate the performance of LLMs on different reasoning tasks by conducting experiments that directly use or draw inspirations from existing datasets on analogical and spatial reasoning. Additionally, to evaluate the ability of LLMs to reason like human, their performance is evaluted on more open-ended, natural language questions. My findings indicate that LLMs excel at analogical and moral reasoning, yet struggle to perform as proficiently on spatial reasoning tasks. I believe these experiments are crucial for informing the future development of LLMs, particularly in contexts that require diverse reasoning proficiencies. By shedding light on the reasoning abilities of LLMs, this study aims to push forward our understanding of how they can better emulate the cognitive abilities of humans.
We report the development of GPT-4, a large-scale, multimodal model which can accept image and text inputs and produce text outputs. While less capable than humans in many real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, including passing a simulated bar exam with a score around the top 10% of test takers. GPT-4 is a Transformer-based model pre-trained to predict the next token in a document. The post-training alignment process results in improved performance on measures of factuality and adherence to desired behavior. A core component of this project was developing infrastructure and optimization methods that behave predictably across a wide range of scales. This allowed us to accurately predict some aspects of GPT-4's performance based on models trained with no more than 1/1,000th the compute of GPT-4.
Building 3D maps of the environment is central to robot navigation, planning, and interaction with objects in a scene. Most existing approaches that integrate semantic concepts with 3D maps largely remain confined to the closed-set setting: they can only reason about a finite set of concepts, pre-defined at training time. Further, these maps can only be queried using class labels, or in recent work, using text prompts. We address both these issues with ConceptFusion, a scene representation that is (1) fundamentally open-set, enabling reasoning beyond a closed set of concepts and (ii) inherently multimodal, enabling a diverse range of possible queries to the 3D map, from language, to images, to audio, to 3D geometry, all working in concert. ConceptFusion leverages the open-set capabilities of today's foundation models pre-trained on internet-scale data to reason about concepts across modalities such as natural language, images, and audio. We demonstrate that pixel-aligned open-set features can be fused into 3D maps via traditional SLAM and multi-view fusion approaches. This enables effective zero-shot spatial reasoning, not needing any additional training or finetuning, and retains long-tailed concepts better than supervised approaches, outperforming them by more than 40% margin on 3D IoU. We extensively evaluate ConceptFusion on a number of real-world datasets, simulated home environments, a real-world tabletop manipulation task, and an autonomous driving platform. We showcase new avenues for blending foundation models with 3D open-set multimodal mapping. For more information, visit our project page https://concept-fusion.github.io or watch our 5-minute explainer video https://www.youtube.com/watch?v=rkXgws8fiDs
Recent large language models (LLMs) have demonstrated remarkable performance on a variety of natural language processing (NLP) tasks, leading to intense excitement about their applicability across various domains. Unfortunately, recent work has also shown that LLMs are unable to perform accurate reasoning nor solve planning problems, which may limit their usefulness for robotics-related tasks. In this work, our central question is whether LLMs are able to translate goals specified in natural language to a structured planning language. If so, LLM can act as a natural interface between the planner and human users; the translated goal can be handed to domain-independent AI planners that are very effective at planning. Our empirical results on GPT 3.5 variants show that LLMs are much better suited towards translation rather than planning. We find that LLMs are able to leverage commonsense knowledge and reasoning to furnish missing details from under-specified goals (as is often the case in natural language). However, our experiments also reveal that LLMs can fail to generate goals in tasks that involve numerical or physical (e.g., spatial) reasoning, and that LLMs are sensitive to the prompts used. As such, these models are promising for translation to structured planning languages, but care should be taken in their use.
Recent times have witnessed an increasing number of applications of deep neural networks towards solving tasks that require superior cognitive abilities, e.g., playing Go, generating art, ChatGPT, etc. Such a dramatic progress raises the question: how generalizable are neural networks in solving problems that demand broad skills? To answer this question, we propose SMART: a Simple Multimodal Algorithmic Reasoning Task and the associated SMART-101 dataset, for evaluating the abstraction, deduction, and generalization abilities of neural networks in solving visuo-linguistic puzzles designed specifically for children in the 6--8 age group. Our dataset consists of 101 unique puzzles; each puzzle comprises a picture and a question, and their solution needs a mix of several elementary skills, including arithmetic, algebra, and spatial reasoning, among others. To scale our dataset towards training deep neural networks, we programmatically generate entirely new instances for each puzzle, while retaining their solution algorithm. To benchmark performances on SMART-101, we propose a vision and language meta-learning model using varied state-of-the-art backbones. Our experiments reveal that while powerful deep models offer reasonable performances on puzzles in a supervised setting, they are not better than random accuracy when analyzed for generalization. We also evaluate the recent ChatGPT and other large language models on a subset of SMART-101 and find that while these models show convincing reasoning abilities, the answers are often incorrect.
Pixel-level labels are particularly expensive to acquire. Hence, pretraining is a critical step to improve models on a task like semantic segmentation. However, prominent algorithms for pretraining neural networks use image-level objectives, e.g. image classification, image-text alignment a la CLIP, or self-supervised contrastive learning. These objectives do not model spatial information, which might be sub-optimal when finetuning on downstream tasks with spatial reasoning. In this work, we pretrain network with a location-aware (LOCA) self-supervised method which fosters the emergence of strong dense features. Specifically, we use both a patch-level clustering scheme to mine dense pseudo-labels and a relative location prediction task to encourage learning about object parts and their spatial arrangements. Our experiments show that LOCA pretraining leads to representations that transfer competitively to challenging and diverse semantic segmentation datasets.
Understanding spatial and visual information is essential for a navigation agent who follows natural language instructions. The current Transformer-based VLN agents entangle the orientation and vision information, which limits the gain from the learning of each information source. In this paper, we design a neural agent with explicit Orientation and Vision modules. Those modules learn to ground spatial information and landmark mentions in the instructions to the visual environment more effectively. To strengthen the spatial reasoning and visual perception of the agent, we design specific pre-training tasks to feed and better utilize the corresponding modules in our final navigation model. We evaluate our approach on both Room2room (R2R) and Room4room (R4R) datasets and achieve the state of the art results on both benchmarks.
Spatial reasoning poses a particular challenge for intelligent agents and is at the same time a prerequisite for their successful interaction and communication in the physical world. One such reasoning task is to describe the position of a target object with respect to the intrinsic orientation of some reference object via relative directions. In this paper, we introduce GRiD-A-3D, a novel diagnostic visual question-answering (VQA) dataset based on abstract objects. Our dataset allows for a fine-grained analysis of end-to-end VQA models' capabilities to ground relative directions. At the same time, model training requires considerably fewer computational resources compared with existing datasets, yet yields a comparable or even higher performance. Along with the new dataset, we provide a thorough evaluation based on two widely known end-to-end VQA architectures trained on GRiD-A-3D. We demonstrate that within a few epochs, the subtasks required to reason over relative directions, such as recognizing and locating objects in a scene and estimating their intrinsic orientations, are learned in the order in which relative directions are intuitively processed.
Many place-related questions can only be answered by complex spatial reasoning, a task poorly supported by factoid question retrieval. Such reasoning using combinations of spatial and non-spatial criteria pertinent to place-related questions is increasingly possible on linked data knowledge bases. Yet, to enable question answering based on linked knowledge bases, natural language questions must first be re-formulated as formal queries. Here, we first present an enhanced version of YAGO2geo, the geospatially-enabled variant of the YAGO2 knowledge base, by linking and adding more than one million places from OpenStreetMap data to YAGO2. We then propose a novel approach to translate the place-related questions into logical representations, theoretically grounded in the core concepts of spatial information. Next, we use a dynamic template-based approach to generate fully executable GeoSPARQL queries from the logical representations. We test our approach using the Geospatial Gold Standard dataset and report substantial improvements over existing methods.
Spatial relations are a basic part of human cognition. However, they are expressed in natural language in a variety of ways, and previous work has suggested that current vision-and-language models (VLMs) struggle to capture relational information. In this paper, we present Visual Spatial Reasoning (VSR), a dataset containing more than 10k natural text-image pairs with 66 types of spatial relations in English (such as: under, in front of, and facing). While using a seemingly simple annotation format, we show how the dataset includes challenging linguistic phenomena, such as varying reference frames. We demonstrate a large gap between human and model performance: the human ceiling is above 95%, while state-of-the-art models only achieve around 70%. We observe that VLMs' by-relation performances have little correlation with the number of training examples and the tested models are in general incapable of recognising relations concerning the orientations of objects.
Inferring spatial relations in natural language is a crucial ability an intelligent system should possess. The bAbI dataset tries to capture tasks relevant to this domain (task 17 and 19). However, these tasks have several limitations. Most importantly, they are limited to fixed expressions, they are limited in the number of reasoning steps required to solve them, and they fail to test the robustness of models to input that contains irrelevant or redundant information. In this paper, we present a new Question-Answering dataset called StepGame for robust multi-hop spatial reasoning in texts. Our experiments demonstrate that state-of-the-art models on the bAbI dataset struggle on the StepGame dataset. Moreover, we propose a Tensor-Product based Memory-Augmented Neural Network (TP-MANN) specialized for spatial reasoning tasks. Experimental results on both datasets show that our model outperforms all the baselines with superior generalization and robustness performance.
Training a referring expression comprehension (ReC) model for a new visual domain requires collecting referring expressions, and potentially corresponding bounding boxes, for images in the domain. While large-scale pre-trained models are useful for image classification across domains, it remains unclear if they can be applied in a zero-shot manner to more complex tasks like ReC. We present ReCLIP, a simple but strong zero-shot baseline that repurposes CLIP, a state-of-the-art large-scale model, for ReC. Motivated by the close connection between ReC and CLIP's contrastive pre-training objective, the first component of ReCLIP is a region-scoring method that isolates object proposals via cropping and blurring, and passes them to CLIP. However, through controlled experiments on a synthetic dataset, we find that CLIP is largely incapable of performing spatial reasoning off-the-shelf. Thus, the second component of ReCLIP is a spatial relation resolver that handles several types of spatial relations. We reduce the gap between zero-shot baselines from prior work and supervised models by as much as 29% on RefCOCOg, and on RefGTA (video game imagery), ReCLIP's relative improvement over supervised ReC models trained on real images is 8%.