Towards Vision Mixture of Experts for Wildlife Monitoring on the Edge

The explosion of IoT sensors in industrial, consumer and remote sensing use cases has come with unprecedented demand for computing infrastructure to transmit and to analyze petabytes of data. Concurrently, the world is slowly shifting its focus towards more sustainable computing. For these reasons, there has been a recent effort to reduce the footprint of related computing infrastructure, especially by deep learning algorithms, for advanced insight generation. The `TinyML' community is actively proposing methods to save communication bandwidth and excessive cloud storage costs while reducing algorithm inference latency and promoting data privacy. Such proposed approaches should ideally process multiple types of data, including time series, audio, satellite images, and video, near the network edge as multiple data streams has been shown to improve the discriminative ability of learning algorithms, especially for generating fine grained results. Incidentally, there has been recent work on data driven conditional computation of subnetworks that has shown real progress in using a single model to share parameters among very different types of inputs such as images and text, reducing the computation requirement of multi-tower multimodal networks. Inspired by such line of work, we explore similar per patch conditional computation for the first time for mobile vision transformers (vision only case), that will eventually be used for single-tower multimodal edge models. We evaluate the model on Cornell Sap Sucker Woods 60, a fine grained bird species discrimination dataset. Our initial experiments uses $4X$ fewer parameters compared to MobileViTV2-1.0 with a $1$% accuracy drop on the iNaturalist '21 birds test data provided as part of the SSW60 dataset.

From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $α$-NeuS

Traditional 3D shape reconstruction techniques from multi-view images, such as structure from motion and multi-view stereo, primarily focus on opaque surfaces. Similarly, recent advances in neural radiance fields and its variants also primarily address opaque objects, encountering difficulties with the complex lighting effects caused by transparent materials. This paper introduces $\alpha$-NeuS, a new method for simultaneously reconstructing thin transparent objects and opaque objects based on neural implicit surfaces (NeuS). Our method leverages the observation that transparent surfaces induce local extreme values in the learned distance fields during neural volumetric rendering, contrasting with opaque surfaces that align with zero level sets. Traditional iso-surfacing algorithms such as marching cubes, which rely on fixed iso-values, are ill-suited for this data. We address this by taking the absolute value of the distance field and developing an optimization method that extracts level sets corresponding to both non-negative local minima and zero iso-values. We prove that the reconstructed surfaces are unbiased for both transparent and opaque objects. To validate our approach, we construct a benchmark that includes both real-world and synthetic scenes, demonstrating its practical utility and effectiveness. Our data and code are publicly available at https://github.com/728388808/alpha-NeuS.

BendVLM: Test-Time Debiasing of Vision-Language Embeddings

Vision-language model (VLM) embeddings have been shown to encode biases present in their training data, such as societal biases that prescribe negative characteristics to members of various racial and gender identities. VLMs are being quickly adopted for a variety of tasks ranging from few-shot classification to text-guided image generation, making debiasing VLM embeddings crucial. Debiasing approaches that fine-tune the VLM often suffer from catastrophic forgetting. On the other hand, fine-tuning-free methods typically utilize a "one-size-fits-all" approach that assumes that correlation with the spurious attribute can be explained using a single linear direction across all possible inputs. In this work, we propose Bend-VLM, a nonlinear, fine-tuning-free approach for VLM embedding debiasing that tailors the debiasing operation to each unique input. This allows for a more flexible debiasing approach. Additionally, we do not require knowledge of the set of inputs a priori to inference time, making our method more appropriate for online, open-set tasks such as retrieval and text guided image generation.

Identifying Implicit Social Biases in Vision-Language Models

Vision-language models, like CLIP (Contrastive Language Image Pretraining), are becoming increasingly popular for a wide range of multimodal retrieval tasks. However, prior work has shown that large language and deep vision models can learn historical biases contained in their training sets, leading to perpetuation of stereotypes and potential downstream harm. In this work, we conduct a systematic analysis of the social biases that are present in CLIP, with a focus on the interaction between image and text modalities. We first propose a taxonomy of social biases called So-B-IT, which contains 374 words categorized across ten types of bias. Each type can lead to societal harm if associated with a particular demographic group. Using this taxonomy, we examine images retrieved by CLIP from a facial image dataset using each word as part of a prompt. We find that CLIP frequently displays undesirable associations between harmful words and specific demographic groups, such as retrieving mostly pictures of Middle Eastern men when asked to retrieve images of a "terrorist". Finally, we conduct an analysis of the source of such biases, by showing that the same harmful stereotypes are also present in a large image-text dataset used to train CLIP models for examples of biases that we find. Our findings highlight the importance of evaluating and addressing bias in vision-language models, and suggest the need for transparency and fairness-aware curation of large pre-training datasets.

ING-VP: MLLMs cannot Play Easy Vision-based Games Yet

As multimodal large language models (MLLMs) continue to demonstrate increasingly competitive performance across a broad spectrum of tasks, more intricate and comprehensive benchmarks have been developed to assess these cutting-edge models. These benchmarks introduce new challenges to core capabilities such as perception, reasoning, and planning. However, existing multimodal benchmarks fall short in providing a focused evaluation of multi-step planning based on spatial relationships in images. To bridge this gap, we present ING-VP, the first INteractive Game-based Vision Planning benchmark, specifically designed to evaluate the spatial imagination and multi-step reasoning abilities of MLLMs. ING-VP features 6 distinct games, encompassing 300 levels, each with 6 unique configurations. A single model engages in over 60,000 rounds of interaction. The benchmark framework allows for multiple comparison settings, including image-text vs. text-only inputs, single-step vs. multi-step reasoning, and with-history vs. without-history conditions, offering valuable insights into the model's capabilities. We evaluated numerous state-of-the-art MLLMs, with the highest-performing model, Claude-3.5 Sonnet, achieving an average accuracy of only 3.37%, far below the anticipated standard. This work aims to provide a specialized evaluation framework to drive advancements in MLLMs' capacity for complex spatial reasoning and planning. The code is publicly available at https://github.com/Thisisus7/ING-VP.git.

CLSP: High-Fidelity Contrastive Language-State Pre-training for Agent State Representation

With the rapid development of artificial intelligence, multimodal learning has become an important research area. For intelligent agents, the state is a crucial modality to convey precise information alongside common modalities like images, videos, and language. This becomes especially clear with the broad adoption of reinforcement learning and multimodal large language models. Nevertheless, the representation of state modality still lags in development. To this end, we propose a High-Fidelity Contrastive Language-State Pre-training (CLSP) method, which can accurately encode state information into general representations for both reinforcement learning and multimodal large language models. Specifically, we first design a pre-training task based on the classification to train an encoder with coarse-grained information. Next, we construct data pairs of states and language descriptions, utilizing the pre-trained encoder to initialize the CLSP encoder. Then, we deploy contrastive learning to train the CLSP encoder to effectively represent precise state information. Additionally, we enhance the representation of numerical information using the Random Fourier Features (RFF) method for high-fidelity mapping. Extensive experiments demonstrate the superior precision and generalization capabilities of our representation, achieving outstanding results in text-state retrieval, reinforcement learning navigation tasks, and multimodal large language model understanding.

PromptTA: Prompt-driven Text Adapter for Source-free Domain Generalization

Source-free domain generalization (SFDG) tackles the challenge of adapting models to unseen target domains without access to source domain data. To deal with this challenging task, recent advances in SFDG have primarily focused on leveraging the text modality of vision-language models such as CLIP. These methods involve developing a transferable linear classifier based on diverse style features extracted from the text and learned prompts or deriving domain-unified text representations from domain banks. However, both style features and domain banks have limitations in capturing comprehensive domain knowledge. In this work, we propose Prompt-Driven Text Adapter (PromptTA) method, which is designed to better capture the distribution of style features and employ resampling to ensure thorough coverage of domain knowledge. To further leverage this rich domain information, we introduce a text adapter that learns from these style features for efficient domain information storage. Extensive experiments conducted on four benchmark datasets demonstrate that PromptTA achieves state-of-the-art performance. The code is available at https://github.com/zhanghr2001/PromptTA.

Single-View 3D Reconstruction via SO(2)-Equivariant Gaussian Sculpting Networks

This paper introduces SO(2)-Equivariant Gaussian Sculpting Networks (GSNs) as an approach for SO(2)-Equivariant 3D object reconstruction from single-view image observations. GSNs take a single observation as input to generate a Gaussian splat representation describing the observed object's geometry and texture. By using a shared feature extractor before decoding Gaussian colors, covariances, positions, and opacities, GSNs achieve extremely high throughput (>150FPS). Experiments demonstrate that GSNs can be trained efficiently using a multi-view rendering loss and are competitive, in quality, with expensive diffusion-based reconstruction algorithms. The GSN model is validated on multiple benchmark experiments. Moreover, we demonstrate the potential for GSNs to be used within a robotic manipulation pipeline for object-centric grasping.

Curriculum Prompting Foundation Models for Medical Image Segmentation

Adapting large pre-trained foundation models, e.g., SAM, for medical image segmentation remains a significant challenge. A crucial step involves the formulation of a series of specialized prompts that incorporate specific clinical instructions. Past works have been heavily reliant on a singular type of prompt for each instance, necessitating manual input of an ideally correct prompt, which is less efficient. To tackle this issue, we propose to utilize prompts of different granularity, which are sourced from original images to provide a broader scope of clinical insights. However, combining prompts of varying types can pose a challenge due to potential conflicts. In response, we have designed a coarse-to-fine mechanism, referred to as curriculum prompting, that progressively integrates prompts of different types. Through extensive experiments on three public medical datasets across various modalities, we demonstrate the effectiveness of our proposed approach, which not only automates the prompt generation process but also yields superior performance compared to other SAM-based medical image segmentation methods. Code is available at: https://github.com/AnnaZzz-zxq/Curriculum-Prompting.

AutoPV: Automatically Design Your Photovoltaic Power Forecasting Model

Photovoltaic power forecasting (PVPF) is a critical area in time series forecasting (TSF), enabling the efficient utilization of solar energy. With advancements in machine learning and deep learning, various models have been applied to PVPF tasks. However, constructing an optimal predictive architecture for specific PVPF tasks remains challenging, as it requires cross-domain knowledge and significant labor costs. To address this challenge, we introduce AutoPV, a novel framework for the automated search and construction of PVPF models based on neural architecture search (NAS) technology. We develop a brand new NAS search space that incorporates various data processing techniques from state-of-the-art (SOTA) TSF models and typical PVPF deep learning models. The effectiveness of AutoPV is evaluated on diverse PVPF tasks using a dataset from the Daqing Photovoltaic Station in China. Experimental results demonstrate that AutoPV can complete the predictive architecture construction process in a relatively short time, and the newly constructed architecture is superior to SOTA predefined models. This work bridges the gap in applying NAS to TSF problems, assisting non-experts and industries in automatically designing effective PVPF models.