Task2Box: Box Embeddings for Modeling Asymmetric Task Relationships

Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery, multi-tasking, and transfer learning. However, many relationships, such as containment and transferability, are naturally asymmetric and current approaches for representation and visualization (e.g., t-SNE) do not readily support this. We propose Task2Box, an approach to represent tasks using box embeddings -- axis-aligned hyperrectangles in low dimensional spaces -- that can capture asymmetric relationships between them through volumetric overlaps. We show that Task2Box accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets, as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g., CLIP, Task2Vec, or attribute based) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations, as well as handcrafted asymmetric distances (e.g., KL divergence). This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face.

Improved Zero-Shot Classification by Adapting VLMs with Text Descriptions

The zero-shot performance of existing vision-language models (VLMs) such as CLIP is limited by the availability of large-scale, aligned image and text datasets in specific domains. In this work, we leverage two complementary sources of information -- descriptions of categories generated by large language models (LLMs) and abundant, fine-grained image classification datasets -- to improve the zero-shot classification performance of VLMs across fine-grained domains. On the technical side, we develop methods to train VLMs with this "bag-level" image-text supervision. We find that simply using these attributes at test-time does not improve performance, but our training strategy, for example, on the iNaturalist dataset, leads to an average improvement of 4-5% in zero-shot classification accuracy for novel categories of birds and flowers. Similar improvements are observed in domains where a subset of the categories was used to fine-tune the model. By prompting LLMs in various ways, we generate descriptions that capture visual appearance, habitat, and geographic regions and pair them with existing attributes such as the taxonomic structure of the categories. We systematically evaluate their ability to improve zero-shot categorization in natural domains. Our findings suggest that geographic priors can be just as effective and are complementary to visual appearance. Our method also outperforms prior work on prompt-based tuning of VLMs. We plan to release the benchmark, consisting of 7 datasets, which will contribute to future research in zero-shot recognition.

Human in-the-Loop Estimation of Cluster Count in Datasets via Similarity-Driven Nested Importance Sampling

Identifying the number of clusters serves as a preliminary goal for many data analysis tasks. A common approach to this problem is to vary the number of clusters in a clustering algorithm (e.g., 'k' in $k$-means) and pick the value that best explains the data. However, the count estimates can be unreliable especially when the image similarity is poor. Human feedback on the pairwise similarity can be used to improve the clustering, but existing approaches do not guarantee accurate count estimates. We propose an approach to produce estimates of the cluster counts in a large dataset given an approximate pairwise similarity. Our framework samples edges guided by the pairwise similarity, and we collect human feedback to construct a statistical estimate of the cluster count. On the technical front we have developed a nested importance sampling approach that yields (asymptotically) unbiased estimates of the cluster count with confidence intervals which can guide human effort. Compared to naive sampling, our similarity-driven sampling produces more accurate estimates of counts and tighter confidence intervals. We evaluate our method on a benchmark of six fine-grained image classification datasets achieving low error rates on the estimated number of clusters with significantly less human labeling effort compared to baselines and alternative active clustering approaches.

PARTICLE: Part Discovery and Contrastive Learning for Fine-grained Recognition

We develop techniques for refining representations for fine-grained classification and segmentation tasks in a self-supervised manner. We find that fine-tuning methods based on instance-discriminative contrastive learning are not as effective, and posit that recognizing part-specific variations is crucial for fine-grained categorization. We present an iterative learning approach that incorporates part-centric equivariance and invariance objectives. First, pixel representations are clustered to discover parts. We analyze the representations from convolutional and vision transformer networks that are best suited for this task. Then, a part-centric learning step aggregates and contrasts representations of parts within an image. We show that this improves the performance on image classification and part segmentation tasks across datasets. For example, under a linear-evaluation scheme, the classification accuracy of a ResNet50 trained on ImageNet using DetCon, a self-supervised learning approach, improves from 35.4% to 42.0% on the Caltech-UCSD Birds, from 35.5% to 44.1% on the FGVC Aircraft, and from 29.7% to 37.4% on the Stanford Cars. We also observe significant gains in few-shot part segmentation tasks using the proposed technique, while instance-discriminative learning was not as effective. Smaller, yet consistent, improvements are also observed for stronger networks based on transformers.

COSE: A Consistency-Sensitivity Metric for Saliency on Image Classification

We present a set of metrics that utilize vision priors to effectively assess the performance of saliency methods on image classification tasks. To understand behavior in deep learning models, many methods provide visual saliency maps emphasizing image regions that most contribute to a model prediction. However, there is limited work on analyzing the reliability of saliency methods in explaining model decisions. We propose the metric COnsistency-SEnsitivity (COSE) that quantifies the equivariant and invariant properties of visual model explanations using simple data augmentations. Through our metrics, we show that although saliency methods are thought to be architecture-independent, most methods could better explain transformer-based models over convolutional-based models. In addition, GradCAM was found to outperform other methods in terms of COSE but was shown to have limitations such as lack of variability for fine-grained datasets. The duality between consistency and sensitivity allow the analysis of saliency methods from different angles. Ultimately, we find that it is important to balance these two metrics for a saliency map to faithfully show model behavior.

LU-NeRF: Scene and Pose Estimation by Synchronizing Local Unposed NeRFs

A critical obstacle preventing NeRF models from being deployed broadly in the wild is their reliance on accurate camera poses. Consequently, there is growing interest in extending NeRF models to jointly optimize camera poses and scene representation, which offers an alternative to off-the-shelf SfM pipelines which have well-understood failure modes. Existing approaches for unposed NeRF operate under limited assumptions, such as a prior pose distribution or coarse pose initialization, making them less effective in a general setting. In this work, we propose a novel approach, LU-NeRF, that jointly estimates camera poses and neural radiance fields with relaxed assumptions on pose configuration. Our approach operates in a local-to-global manner, where we first optimize over local subsets of the data, dubbed mini-scenes. LU-NeRF estimates local pose and geometry for this challenging few-shot task. The mini-scene poses are brought into a global reference frame through a robust pose synchronization step, where a final global optimization of pose and scene can be performed. We show our LU-NeRF pipeline outperforms prior attempts at unposed NeRF without making restrictive assumptions on the pose prior. This allows us to operate in the general SE(3) pose setting, unlike the baselines. Our results also indicate our model can be complementary to feature-based SfM pipelines as it compares favorably to COLMAP on low-texture and low-resolution images.

DISCount: Counting in Large Image Collections with Detector-Based Importance Sampling

Many modern applications use computer vision to detect and count objects in massive image collections. However, when the detection task is very difficult or in the presence of domain shifts, the counts may be inaccurate even with significant investments in training data and model development. We propose DISCount -- a detector-based importance sampling framework for counting in large image collections that integrates an imperfect detector with human-in-the-loop screening to produce unbiased estimates of counts. We propose techniques for solving counting problems over multiple spatial or temporal regions using a small number of screened samples and estimate confidence intervals. This enables end-users to stop screening when estimates are sufficiently accurate, which is often the goal in a scientific study. On the technical side we develop variance reduction techniques based on control variates and prove the (conditional) unbiasedness of the estimators. DISCount leads to a 9-12x reduction in the labeling costs over naive screening for tasks we consider, such as counting birds in radar imagery or estimating damaged buildings in satellite imagery, and also surpasses alternative covariate-based screening approaches in efficiency.

Accidental Turntables: Learning 3D Pose by Watching Objects Turn

We propose a technique for learning single-view 3D object pose estimation models by utilizing a new source of data -- in-the-wild videos where objects turn. Such videos are prevalent in practice (e.g., cars in roundabouts, airplanes near runways) and easy to collect. We show that classical structure-from-motion algorithms, coupled with the recent advances in instance detection and feature matching, provides surprisingly accurate relative 3D pose estimation on such videos. We propose a multi-stage training scheme that first learns a canonical pose across a collection of videos and then supervises a model for single-view pose estimation. The proposed technique achieves competitive performance with respect to existing state-of-the-art on standard benchmarks for 3D pose estimation, without requiring any pose labels during training. We also contribute an Accidental Turntables Dataset, containing a challenging set of 41,212 images of cars in cluttered backgrounds, motion blur and illumination changes that serves as a benchmark for 3D pose estimation.

MvDeCor: Multi-view Dense Correspondence Learning for Fine-grained 3D Segmentation

We propose to utilize self-supervised techniques in the 2D domain for fine-grained 3D shape segmentation tasks. This is inspired by the observation that view-based surface representations are more effective at modeling high-resolution surface details and texture than their 3D counterparts based on point clouds or voxel occupancy. Specifically, given a 3D shape, we render it from multiple views, and set up a dense correspondence learning task within the contrastive learning framework. As a result, the learned 2D representations are view-invariant and geometrically consistent, leading to better generalization when trained on a limited number of labeled shapes compared to alternatives that utilize self-supervision in 2D or 3D alone. Experiments on textured (RenderPeople) and untextured (PartNet) 3D datasets show that our method outperforms state-of-the-art alternatives in fine-grained part segmentation. The improvements over baselines are greater when only a sparse set of views is available for training or when shapes are textured, indicating that MvDeCor benefits from both 2D processing and 3D geometric reasoning.

Cross-Modal 3D Shape Generation and Manipulation

Creating and editing the shape and color of 3D objects require tremendous human effort and expertise. Compared to direct manipulation in 3D interfaces, 2D interactions such as sketches and scribbles are usually much more natural and intuitive for the users. In this paper, we propose a generic multi-modal generative model that couples the 2D modalities and implicit 3D representations through shared latent spaces. With the proposed model, versatile 3D generation and manipulation are enabled by simply propagating the editing from a specific 2D controlling modality through the latent spaces. For example, editing the 3D shape by drawing a sketch, re-colorizing the 3D surface via painting color scribbles on the 2D rendering, or generating 3D shapes of a certain category given one or a few reference images. Unlike prior works, our model does not require re-training or fine-tuning per editing task and is also conceptually simple, easy to implement, robust to input domain shifts, and flexible to diverse reconstruction on partial 2D inputs. We evaluate our framework on two representative 2D modalities of grayscale line sketches and rendered color images, and demonstrate that our method enables various shape manipulation and generation tasks with these 2D modalities.