DMTG: One-Shot Differentiable Multi-Task Grouping

We aim to address Multi-Task Learning (MTL) with a large number of tasks by Multi-Task Grouping (MTG). Given N tasks, we propose to simultaneously identify the best task groups from 2^N candidates and train the model weights simultaneously in one-shot, with the high-order task-affinity fully exploited. This is distinct from the pioneering methods which sequentially identify the groups and train the model weights, where the group identification often relies on heuristics. As a result, our method not only improves the training efficiency, but also mitigates the objective bias introduced by the sequential procedures that potentially lead to a suboptimal solution. Specifically, we formulate MTG as a fully differentiable pruning problem on an adaptive network architecture determined by an underlying Categorical distribution. To categorize N tasks into K groups (represented by K encoder branches), we initially set up KN task heads, where each branch connects to all N task heads to exploit the high-order task-affinity. Then, we gradually prune the KN heads down to N by learning a relaxed differentiable Categorical distribution, ensuring that each task is exclusively and uniquely categorized into only one branch. Extensive experiments on CelebA and Taskonomy datasets with detailed ablations show the promising performance and efficiency of our method. The codes are available at https://github.com/ethanygao/DMTG.

Towards Human-Level 3D Relative Pose Estimation: Generalizable, Training-Free, with Single Reference

Humans can easily deduce the relative pose of an unseen object, without label/training, given only a single query-reference image pair. This is arguably achieved by incorporating (i) 3D/2.5D shape perception from a single image, (ii) render-and-compare simulation, and (iii) rich semantic cue awareness to furnish (coarse) reference-query correspondence. Existing methods implement (i) by a 3D CAD model or well-calibrated multiple images and (ii) by training a network on specific objects, which necessitate laborious ground-truth labeling and tedious training, potentially leading to challenges in generalization. Moreover, (iii) was less exploited in the paradigm of (ii), despite that the coarse correspondence from (iii) enhances the compare process by filtering out non-overlapped parts under substantial pose differences/occlusions. Motivated by this, we propose a novel 3D generalizable relative pose estimation method by elaborating (i) with a 2.5D shape from an RGB-D reference, (ii) with an off-the-shelf differentiable renderer, and (iii) with semantic cues from a pretrained model like DINOv2. Specifically, our differentiable renderer takes the 2.5D rotatable mesh textured by the RGB and the semantic maps (obtained by DINOv2 from the RGB input), then renders new RGB and semantic maps (with back-surface culling) under a novel rotated view. The refinement loss comes from comparing the rendered RGB and semantic maps with the query ones, back-propagating the gradients through the differentiable renderer to refine the 3D relative pose. As a result, our method can be readily applied to unseen objects, given only a single RGB-D reference, without label/training. Extensive experiments on LineMOD, LM-O, and YCB-V show that our training-free method significantly outperforms the SOTA supervised methods, especially under the rigorous Acc@5/10/15{\deg} metrics and the challenging cross-dataset settings.

Optimization-based Structural Pruning for Large Language Models without Back-Propagation

Compared to the moderate size of neural network models, structural weight pruning on the Large-Language Models (LLMs) imposes a novel challenge on the efficiency of the pruning algorithms, due to the heavy computation/memory demands of the LLMs. Recent efficient LLM pruning methods typically operate at the post-training phase without the expensive weight finetuning, however, their pruning criteria often rely on heuristically designed metrics, potentially leading to suboptimal performance. We instead propose a novel optimization-based structural pruning that learns the pruning masks in a probabilistic space directly by optimizing the loss of the pruned model. To preserve the efficiency, our method 1) works at post-training phase} and 2) eliminates the back-propagation through the LLM per se during the optimization (i.e., only requires the forward pass of the LLM). We achieve this by learning an underlying Bernoulli distribution to sample binary pruning masks, where we decouple the Bernoulli parameters from the LLM loss, thus facilitating an efficient optimization via a policy gradient estimator without back-propagation. As a result, our method is able to 1) operate at structural granularities of channels, heads, and layers, 2) support global and heterogeneous pruning (i.e., our method automatically determines different redundancy for different layers), and 3) optionally use a metric-based method as initialization (of our Bernoulli distributions). Extensive experiments on LLaMA, LLaMA-2, and Vicuna using the C4 and WikiText2 datasets demonstrate that our method operates for 2.7 hours with around 35GB memory for the 13B models on a single A100 GPU, and our pruned models outperform the state-of-the-arts w.r.t. perplexity. Codes will be released.

Aux-NAS: Exploiting Auxiliary Labels with Negligibly Extra Inference Cost

We aim at exploiting additional auxiliary labels from an independent (auxiliary) task to boost the primary task performance which we focus on, while preserving a single task inference cost of the primary task. While most existing auxiliary learning methods are optimization-based relying on loss weights/gradients manipulation, our method is architecture-based with a flexible asymmetric structure for the primary and auxiliary tasks, which produces different networks for training and inference. Specifically, starting from two single task networks/branches (each representing a task), we propose a novel method with evolving networks where only primary-to-auxiliary links exist as the cross-task connections after convergence. These connections can be removed during the primary task inference, resulting in a single-task inference cost. We achieve this by formulating a Neural Architecture Search (NAS) problem, where we initialize bi-directional connections in the search space and guide the NAS optimization converging to an architecture with only the single-side primary-to-auxiliary connections. Moreover, our method can be incorporated with optimization-based auxiliary learning approaches. Extensive experiments with six tasks on NYU v2, CityScapes, and Taskonomy datasets using VGG, ResNet, and ViT backbones validate the promising performance. The codes are available at https://github.com/ethanygao/Aux-NAS.

Dual Relation Mining Network for Zero-Shot Learning

Zero-shot learning (ZSL) aims to recognize novel classes through transferring shared semantic knowledge (e.g., attributes) from seen classes to unseen classes. Recently, attention-based methods have exhibited significant progress which align visual features and attributes via a spatial attention mechanism. However, these methods only explore visual-semantic relationship in the spatial dimension, which can lead to classification ambiguity when different attributes share similar attention regions, and semantic relationship between attributes is rarely discussed. To alleviate the above problems, we propose a Dual Relation Mining Network (DRMN) to enable more effective visual-semantic interactions and learn semantic relationship among attributes for knowledge transfer. Specifically, we introduce a Dual Attention Block (DAB) for visual-semantic relationship mining, which enriches visual information by multi-level feature fusion and conducts spatial attention for visual to semantic embedding. Moreover, an attribute-guided channel attention is utilized to decouple entangled semantic features. For semantic relationship modeling, we utilize a Semantic Interaction Transformer (SIT) to enhance the generalization of attribute representations among images. Additionally, a global classification branch is introduced as a complement to human-defined semantic attributes, and we then combine the results with attribute-based classification. Extensive experiments demonstrate that the proposed DRMN leads to new state-of-the-art performances on three standard ZSL benchmarks, i.e., CUB, SUN, and AwA2.

Anchor-based Robust Finetuning of Vision-Language Models

We aim at finetuning a vision-language model without hurting its out-of-distribution (OOD) generalization. We address two types of OOD generalization, i.e., i) domain shift such as natural to sketch images, and ii) zero-shot capability to recognize the category that was not contained in the finetune data. Arguably, the diminished OOD generalization after finetuning stems from the excessively simplified finetuning target, which only provides the class information, such as ``a photo of a [CLASS]''. This is distinct from the process in that CLIP was pretrained, where there is abundant text supervision with rich semantic information. Therefore, we propose to compensate for the finetune process using auxiliary supervision with rich semantic information, which acts as anchors to preserve the OOD generalization. Specifically, two types of anchors are elaborated in our method, including i) text-compensated anchor which uses the images from the finetune set but enriches the text supervision from a pretrained captioner, ii) image-text-pair anchor which is retrieved from the dataset similar to pretraining data of CLIP according to the downstream task, associating with the original CLIP text with rich semantics. Those anchors are utilized as auxiliary semantic information to maintain the original feature space of CLIP, thereby preserving the OOD generalization capabilities. Comprehensive experiments demonstrate that our method achieves in-distribution performance akin to conventional finetuning while attaining new state-of-the-art results on domain shift and zero-shot learning benchmarks.

3D Building Reconstruction from Monocular Remote Sensing Images with Multi-level Supervisions

3D building reconstruction from monocular remote sensing images is an important and challenging research problem that has received increasing attention in recent years, owing to its low cost of data acquisition and availability for large-scale applications. However, existing methods rely on expensive 3D-annotated samples for fully-supervised training, restricting their application to large-scale cross-city scenarios. In this work, we propose MLS-BRN, a multi-level supervised building reconstruction network that can flexibly utilize training samples with different annotation levels to achieve better reconstruction results in an end-to-end manner. To alleviate the demand on full 3D supervision, we design two new modules, Pseudo Building Bbox Calculator and Roof-Offset guided Footprint Extractor, as well as new tasks and training strategies for different types of samples. Experimental results on several public and new datasets demonstrate that our proposed MLS-BRN achieves competitive performance using much fewer 3D-annotated samples, and significantly improves the footprint extraction and 3D reconstruction performance compared with current state-of-the-art. The code and datasets of this work will be released at https://github.com/opendatalab/MLS-BRN.git.

H2RSVLM: Towards Helpful and Honest Remote Sensing Large Vision Language Model

The generic large Vision-Language Models (VLMs) is rapidly developing, but still perform poorly in Remote Sensing (RS) domain, which is due to the unique and specialized nature of RS imagery and the comparatively limited spatial perception of current VLMs. Existing Remote Sensing specific Vision Language Models (RSVLMs) still have considerable potential for improvement, primarily owing to the lack of large-scale, high-quality RS vision-language datasets. We constructed HqDC-1.4M, the large scale High quality and Detailed Captions for RS images, containing 1.4 million image-caption pairs, which not only enhance the RSVLM's understanding of RS images but also significantly improve the model's spatial perception abilities, such as localization and counting, thereby increasing the helpfulness of the RSVLM. Moreover, to address the inevitable "hallucination" problem in RSVLM, we developed RSSA, the first dataset aimed at enhancing the Self-Awareness capability of RSVLMs. By incorporating a variety of unanswerable questions into typical RS visual question-answering tasks, RSSA effectively improves the truthfulness and reduces the hallucinations of the model's outputs, thereby enhancing the honesty of the RSVLM. Based on these datasets, we proposed the H2RSVLM, the Helpful and Honest Remote Sensing Vision Language Model. H2RSVLM has achieved outstanding performance on multiple RS public datasets and is capable of recognizing and refusing to answer the unanswerable questions, effectively mitigating the incorrect generations. We will release the code, data and model weights at https://github.com/opendatalab/H2RSVLM .

Unleashing Unlabeled Data: A Paradigm for Cross-View Geo-Localization

This paper investigates the effective utilization of unlabeled data for large-area cross-view geo-localization (CVGL), encompassing both unsupervised and semi-supervised settings. Common approaches to CVGL rely on ground-satellite image pairs and employ label-driven supervised training. However, the cost of collecting precise cross-view image pairs hinders the deployment of CVGL in real-life scenarios. Without the pairs, CVGL will be more challenging to handle the significant imaging and spatial gaps between ground and satellite images. To this end, we propose an unsupervised framework including a cross-view projection to guide the model for retrieving initial pseudo-labels and a fast re-ranking mechanism to refine the pseudo-labels by leveraging the fact that ``the perfectly paired ground-satellite image is located in a unique and identical scene". The framework exhibits competitive performance compared with supervised works on three open-source benchmarks. Our code and models will be released on https://github.com/liguopeng0923/UCVGL.

Learning Cross-view Visual Geo-localization without Ground Truth

Cross-View Geo-Localization (CVGL) involves determining the geographical location of a query image by matching it with a corresponding GPS-tagged reference image. Current state-of-the-art methods predominantly rely on training models with labeled paired images, incurring substantial annotation costs and training burdens. In this study, we investigate the adaptation of frozen models for CVGL without requiring ground truth pair labels. We observe that training on unlabeled cross-view images presents significant challenges, including the need to establish relationships within unlabeled data and reconcile view discrepancies between uncertain queries and references. To address these challenges, we propose a self-supervised learning framework to train a learnable adapter for a frozen Foundation Model (FM). This adapter is designed to map feature distributions from diverse views into a uniform space using unlabeled data exclusively. To establish relationships within unlabeled data, we introduce an Expectation-Maximization-based Pseudo-labeling module, which iteratively estimates associations between cross-view features and optimizes the adapter. To maintain the robustness of the FM's representation, we incorporate an information consistency module with a reconstruction loss, ensuring that adapted features retain strong discriminative ability across views. Experimental results demonstrate that our proposed method achieves significant improvements over vanilla FMs and competitive accuracy compared to supervised methods, while necessitating fewer training parameters and relying solely on unlabeled data. Evaluation of our adaptation for task-specific models further highlights its broad applicability.