We introduce VideoComp, a benchmark and learning framework for advancing video-text compositionality understanding, aimed at improving vision-language models (VLMs) in fine-grained temporal alignment. Unlike existing benchmarks focused on static image-text compositionality or isolated single-event videos, our benchmark targets alignment in continuous multi-event videos. Leveraging video-text datasets with temporally localized event captions (e.g. ActivityNet-Captions, YouCook2), we construct two compositional benchmarks, ActivityNet-Comp and YouCook2-Comp. We create challenging negative samples with subtle temporal disruptions such as reordering, action word replacement, partial captioning, and combined disruptions. These benchmarks comprehensively test models’ compositional sensitivity across extended, cohesive video-text sequences. To improve model performance, we propose a hierarchical pairwise preference loss that strengthens alignment with temporally accurate pairs and gradually penalizes increasingly disrupted ones, encouraging fine-grained compositional learning. To mitigate the limited availability of densely annotated video data, we introduce a pretraining strategy that concatenates short video-caption pairs to simulate multi-event sequences. We evaluate video-text foundational models and large multimodal models (LMMs) on our benchmark, identifying both strengths and areas for improvement in compositionality. Overall, our work provides a comprehensive framework for evaluating and enhancing model capabilities in achieving fine-grained, temporally coherent video-text alignment.

We explore the boundaries of scaling up a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. Our model advances the state-of-the-art on most vision-and-language benchmarks considered (20+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.

One of the main challenges of multimodal learning is combining multiple heterogeneous modalities, e.g., video, audio, and text. Video and audio are obtained at much higher rates than text and are roughly aligned in time. They are often not synchronized with text, which comes as a global context, e.g. a title, or a description. Furthermore, video and audio inputs are of much larger volumes, and grow as the video length increases, which naturally requires more compute dedicated to these modalities, and makes modeling of long-range dependencies harder. We here decouple the multimodal modeling, dividing it into separate autoregressive models, processing the inputs according to the characteristics of the modalities. We propose a multimodal model, consisting of an autoregressive component for the time-synchronized modalities (audio and video), and an autoregressive component for the context modalities which are not necessarily aligned in time but are still sequential. To address the long-sequences of the video-audio inputs, we further partition the video and audio sequences in consecutive snippets and autoregressively process their representations. To that end, we propose a Combiner mechanism, which models the audio-video information jointly, producing compact but expressive representations. This allows us to scale to 512 input video frames without increase in model parameters. Our approach achieves the state-of-the-art on multiple well established multimodal benchmarks. It effectively addresses the high computational demand of media inputs by learning compact representations, controlling the sequence length of the audio-video feature representations, and modeling their dependencies in time.

We present Region-aware Open-vocabulary Vision Transformers (RO-ViT) - a contrastive image-text pretraining recipe to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we propose to randomly crop and resize regions of positional embeddings instead of using the whole image positional embeddings. This better matches the use of positional embeddings at region-level in the detection finetuning phase. In addition, we replace the common softmax cross entropy loss in contrastive learning with focal loss to better learn the informative yet difficult examples. Finally, we leverage recent advances in novel object proposals to improve open-vocabulary detection finetuning. We evaluate our full model on the LVIS and COCO open-vocabulary detection benchmarks and zero-shot transfer. RO-ViT achieves a state-of-the-art 32.1 APron LVIS, surpassing the best existing approach by +5.8 points in addition to competitive zero-shot transfer detection. Surprisingly, RO-ViT improves the image-level representation as well and achieves the state of the art on 9 out of 12 metrics on COCO and Flickr image-text retrieval benchmarks, outperforming competitive approaches with larger models.

We present a simple approach which can turn a ViT en-coder into an efficient video model, which can seamlessly work with both image and video inputs. By sparsely sam-pling the inputs, the model is able to do training and in-ference from both input modalities. The model is easily scalable and can be adapted to large-scale pre-trained ViTs without requiring full finetuning. The model achieves SOTA results11https://sites.google.com/view/tubevit.

We present Contrastive Feature Masking Vision Trans-former (CFM-ViT) - an image-text pretraining methodology that achieves simultaneous learning of image- and region-level representation for open-vocabulary object detection (OVD). Our approach combines the masked autoencoder (MAE) objective into the contrastive learning objective to improve the representation for localization tasks. Un-like standard MAE, we perform reconstruction in the joint image-text embedding space, rather than the pixel space as is customary with the classical MAE method, which causes the model to better learn region-level semantics. Moreover, we introduce Positional Embedding Dropout (PED) to address scale variation between image-text pretraining and detection finetuning by randomly dropping out the positional embeddings during pretraining. PED improves detection performance and enables the use of a frozen ViT backbone as a region classifier, preventing the forgetting of open-vocabulary knowledge during detection finetuning. On LVIS open-vocabulary detection benchmark, CFM-ViT achieves a state-of-the-art 33.9 APr, surpassing the best approach by 7.6 points and achieves better zero-shot detection transfer. Finally, CFM-ViT acquires strong image-level representation, outperforming the state of the art on 8 out of 12 metrics on zero-shot image-text retrieval benchmarks.

Shelves are common in homes, warehouses, and commercial settings due to their storage efficiency. However, this efficiency comes at the cost of reduced visibility and accessibility. When looking from a side (lateral) view of a shelf, most objects will be fully occluded, resulting in a constrained lateral-access mechanical search problem. To address this problem, we introduce: (1) a novel bluction tool, which combines a thin pushing blade and a suction cup gripper, (2) a simulation pipeline and perception model that combine ray-casting with 2D Minkowski sums to efficiently generate target occupancy distributions, and (3) a novel search policy, which optimally reduces target object distribution support area using the bluction tool. Experimental data from 2000 simulated shelf trials and 18 trials with a physical Fetch robot suggest that a bluction tool can improve the average success rate by 26% in simulation and 67% in physical experiments over the highest-performing push-only policy.

Many computer vision tasks address the problem of scene understanding and are naturally interrelated e.g. object classification, detection, scene segmentation, depth estimation, etc. We show that we can leverage the inherent relationships among collections of tasks, as they are trained jointly, supervising each other through their known relationships via consistency losses. Furthermore, explicitly utilizing the relationships between tasks allows improving their performance while dramatically reducing the need for labeled data, and allows training with additional unsupervised or simulated data. We demonstrate a distributed joint training algorithm with task-level parallelism, which affords a high degree of asynchronicity and robustness. This allows learning across multiple tasks, or with large amounts of input data, at scale. We demonstrate our framework on subsets of the following collection of tasks: depth and normal prediction, semantic segmentation, 3D motion and egomotion estimation, and object tracking and 3D detection in point clouds. We observe improved performance across these tasks, especially in the low-label regime.

3D perception of object shapes from RGB image input is fundamental towards semantic scene understanding, grounding image-based perception in our spatially 3dimensional real-world environments. To achieve a mapping between image views of objects and 3D shapes, we leverage CAD model priors from existing large-scale databases, and propose a novel approach towards constructing a joint embedding space between 2D images and 3D CAD models in a patch-wise fashion – establishing correspondences between patches of an image view of an object and patches of CAD geometry. This enables part similarity reasoning for retrieving similar CADs to a new image view without exact matches in the database. Our patch embedding provides more robust CAD retrieval for shape estimation in our end-to-end estimation of CAD model shape and pose for detected objects in a single input image. Experiments on in-the-wild, complex imagery from ScanNet show that our approach is more robust than state of the art in real-world scenarios without any exact CAD matches.

We present 4D-Net, a 3D object detection approach, which utilizes 3D Point Cloud and RGB sensing information, both in time. We are able to incorporate the 4D information by performing a novel dynamic connection learning across various feature representations and levels of abstraction, as well as by observing geometric constraints. Our approach outperforms the state-of-the-art and strong base-lines on the Waymo Open Dataset. 4D-Net is better able to use motion cues and dense image information to detect distant objects more successfully. We will open source the code.

We propose a vision-based architecture search algorithm for robot manipulation learning, which discovers interactions between low dimension action inputs and high dimensional visual inputs. Our approach automatically designs architectures while training on the task – discovering novel ways of combining and attending image feature representations with actions as well as features from previous layers. The obtained new architectures demonstrate better task success rates, in some cases with a large margin, compared to a recent high performing baseline. Our real robot experiments also confirm that it improves grasping performance by 6%. This is the first approach to demonstrate a successful neural architecture search and attention connectivity search for a real-robot task.

Finding an occluded object in a lateral access environment such as a shelf or cabinet is a problem that arises in many contexts such as warehouses, retail, healthcare, shipping, and homes. While this problem, known as mechanical search, is well-studied in overhead access environments, lateral access environments introduce constraints on the poses of objects and on available grasp actions, and pushing actions are preferred to preserve the environment structure. We propose LAX-RAY (Lateral Access maXimal Reduction in support Area of occupancY distribution): a system that combines target object occupancy distribution prediction with a mechanical search policy that sequentially pushes occluding objects to reveal a given target object. For scenarios with extruded polygonal objects, we introduce two lateral-access search policies that encode a history of predicted target distributions and can plan up to three actions into the future. We introduce a First-Order Shelf Simulator (FOSS) and use it to evaluate these policies in 800 simulated random shelf environments per policy. We also evaluate in 5 physical shelf environments using a Fetch robot with an embedded PrimeSense RGBD Camera and an attached pushing blade. The policies outperform baselines by up to 25% in simulation and up to 60% in physical experiments. Additionally, the two-step prediction policy is the highest performing in simulation for 8 objects with a 69% success rate, suggesting a tradeoff between future information and prediction errors. Code, videos, and supplementary material can be found at https://sites.google.com/berkeley.edu/lax-ray.

In this paper, we introduce a novel visual representation learning which relies on a handful of adaptively learned tokens, and which is applicable to both image and video understanding tasks. Instead of relying on hand-designed splitting strategies to obtain visual tokens and processing a large number of densely sampled patches for attention, our approach learns to mine important tokens in visual data. This results in efficiently and effectively finding a few important visual tokens and enables modeling of pairwise attention between such tokens, over a longer temporal horizon for videos, or the spatial content in image frames. Our experiments demonstrate strong performance on several challenging benchmarks for video recognition tasks. Importantly, due to our tokens being adaptive, we accomplish competitive results at significantly reduced computational cost. We establish new state-of-the-arts on multiple video datasets, including Kinetics-400, Kinetics-600, Charades, and AViD.

Estimating the 3D pose of desktop objects is crucial for applications such as robotic manipulation. Many existing approaches to this problem require a depth map of the object for both training and prediction, which restricts them to opaque, lambertian objects that produce good returns in an RGBD sensor. In this paper we forgo using a depth sensor in favor of raw stereo input. We address two problems: first, we establish an easy method for capturing and labeling 3D keypoints on desktop objects with an RGB camera; and second, we develop a deep neural network, called KeyPose, that learns to accurately predict object poses using 3D keypoints, from stereo input, and works even for transparent objects. To evaluate the performance of our method, we create a dataset of 15 clear objects in five classes, with 48K 3D-keypoint labeled images. We train both instance and category models, and show generalization to new textures, poses, and objects. KeyPose surpasses state-of-the-art performance in 3D pose estimation on this dataset by factors of 1.5 to 3.5, even in cases where the competing method is provided with ground-truth depth. Stereo input is essential for this performance as it improves results compared to using monocular input by a factor of 2. We will release a public version of the data capture and labeling pipeline, the transparent object database, and the KeyPose models and evaluation code. Project website: https://sites.google.com/corp/view/keypose.

We present a new method to learn video representations from large-scale unlabeled video data. Ideally, this representation will be generic and transferable, directly usable for new tasks such as action recognition and zero or few-shot learning. We formulate unsupervised representation learning as a multi-modal, multi-task learning problem, where the representations are shared across different modalities via distillation. Further, we introduce the concept of loss function evolution by using an evolutionary search algorithm to automatically find optimal combination of loss functions capturing many (self-supervised) tasks and modalities. Thirdly, we propose an unsupervised representation evaluation metric using distribution matching to a large unlabeled dataset as a prior constraint, based on Zipf's law. This unsupervised constraint, which is not guided by any labeling, produces similar results to weakly-supervised, task-specific ones. The proposed unsupervised representation learning results in a single RGB network and outperforms previous methods. Notably, it is also more effective than several label-based methods (e.g., ImageNet), with the exception of large, fully labeled video datasets.

Mapping and localization, preferably from a small number of observations, are fundamental tasks in robotics. We address these tasks by combining spatial structure (differentiable mapping) and end-to-end learning in a novel neural network architecture: the Differentiable Mapping Network (DMN). The DMN constructs a spatially structured view-embedding map and uses it for subsequent visual localization with a particle filter. Since the DMN architecture is end-to-end differentiable, we can jointly learn the map representation and localization using gradient descent. We apply the DMN to sparse visual localization, where a robot needs to localize in a new environment with respect to a small number of images from known viewpoints. We evaluate the DMN using simulated environments and a challenging real-world Street View dataset. We find that the DMN learns effective map representations for visual localization. The benefit of spatial structure increases with larger environments, more viewpoints for mapping, and when training data is scarce. Project website: https://sites.google.com/view/differentiable-mapping.

For applications in e-commerce, warehouses, healthcare, and home service, robots are often required to search through heaps of objects to grasp a specific target object. For mechanical search, we introduce X-Ray, an algorithm based on learned occupancy distributions. We train a neural network using a synthetic dataset of RGBD heap images labeled for a set of standard bounding box targets with varying aspect ratios. X-Ray minimizes support of the learned distribution as part of a mechanical search policy in both simulated and real environments. We benchmark these policies against two baseline policies on 1,000 heaps of 15 objects in simulation where the target object is partially or fully occluded. Results suggest that X-Ray is significantly more efficient, as it succeeds in extracting the target object 82% of the time, 15% more often than the best-performing baseline. Experiments on an ABB YuMi robot with 20 heaps of 25 household objects suggest that the learned policy transfers easily to a physical system, where it outperforms baseline policies by 15% in success rate with 17% fewer actions. Datasets, videos, and experiments are available at https://sites.google.com/berkeley.edu/x-ray.

We introduce Probabilistic Object Detection, the task of detecting objects in images and accurately quantifying the spatial and semantic uncertainties of the detections. Given the lack of methods capable of assessing such probabilistic object detections, we present the new Probability-based Detection Quality measure (PDQ). Unlike AP-based measures, PDQ has no arbitrary thresholds and rewards spatial and label quality, and foreground/background separation quality while explicitly penalising false positive and false negative detections. We contrast PDQ with existing mAP and moLRP measures by evaluating state-of-the-art detectors and a Bayesian object detector based on Monte Carlo Dropout. Our experiments indicate that conventional object detectors tend to be spatially overconfident and thus perform poorly on the task of probabilistic object detection. Our paper aims to encourage the development of new object detection approaches that provide detections with accurately estimated spatial and label uncertainties and are of critical importance for deployment on robots and embodied AI systems in the real world.

Learning to predict scene depth from RGB inputs is a challenging task both for indoor and outdoor robot navigation. In this work we address unsupervised learning of scene depth and robot ego-motion where supervision is provided by monocular videos, as cameras are the cheapest, least restrictive and most ubiquitous sensor for robotics. Previous work in unsupervised image-to-depth learning has established strong baselines in the domain. We propose a novel approach which produces higher quality results, is able to model moving objects and is shown to transfer across data domains, e.g. from outdoors to indoor scenes. The main idea is to introduce geometric structure in the learning process, by modeling the scene and the individual objects; camera ego-motion and object motions are learned from monocular videos as input. Furthermore an online refinement method is introduced to adapt learning on the fly to unknown domains. The proposed approach outperforms all state-of-the-art approaches, including those that handle motion e.g. through learned flow. Our results are comparable in quality to the ones which used stereo as supervision and significantly improve depth prediction on scenes and datasets which contain a lot of object motion. The approach is of practical relevance, as it allows transfer across environments, by transferring models trained on data collected for robot navigation in urban scenes to indoor navigation settings. The code associated with this paper can be found at https://sites.google.com/view/struct2depth.

We present a novel method for simultaneous learning of depth, egomotion, object motion, and camera intrinsics from monocular videos, using only consistency across neighboring video frames as supervision signal. Similarly to prior work, our method learns by applying differentiable warping to frames and comparing the result to adjacent ones, but it provides several improvements: We address occlusions geometrically and differentiably, directly using the depth maps as predicted during training. We introduce randomized layer normalization, a novel powerful regularizer, and we account for object motion relative to the scene. To the best of our knowledge, our work is the first to learn the camera intrinsic parameters, including lens distortion, from video in an unsupervised manner, thereby allowing us to extract accurate depth and motion from arbitrary videos of unknown origin at scale. We evaluate our results on the Cityscapes, KITTI and EuRoC datasets, establishing new state of the art on depth prediction and odometry, and demonstrate qualitatively that depth prediction can be learned from a collection of YouTube videos. The code will be open sourced once anonymity is lifted.

Instance segmentation aims to detect and segment individual objects in a scene. Most existing methods rely on precise mask annotations of every category. However, it is difficult and costly to segment objects in novel categories because a large number of mask annotations is required. We introduce ShapeMask, which learns the intermediate concept of object shape to address the problem of generalization in instance segmentation to novel categories. ShapeMask starts with a bounding box detection and gradually refines it by first estimating the shape of the detected object through a collection of shape priors. Next, ShapeMask refines the coarse shape into an instance level mask by learning instance embeddings. The shape priors provide a strong cue for object-like prediction, and the instance embeddings model the instance specific appearance information. ShapeMask significantly outperforms the state-of-the-art by 6.4 and 3.8 AP when learning across categories, and obtains competitive performance in the fully supervised setting. It is also robust to inaccurate detections, decreased model capacity, and small training data. Moreover, it runs efficiently with 150ms inference time on a GPU and trains within 11 hours on TPUs. With a larger backbone model, ShapeMask increases the gap with state-of-the-art to 9.4 and 6.2 AP across categories. Code will be publicly available at: https://sites.google.com/view/shapemask/home.

We present a new method for finding video CNN architectures that more optimally capture rich spatio-temporal information in videos. Previous work, taking advantage of 3D convolutions, obtained promising results by manually designing CNN video architectures. We here develop a novel evolutionary algorithm that automatically explores models with different types and combinations of layers to jointly learn interactions between spatial and temporal aspects of video representations. We demonstrate the generality of this algorithm by applying it to two meta-architectures. Further, we propose a new component, the iTGM layer, which more efficiently utilizes its parameters to allow learning of space-time interactions over longer time horizons. The iTGM layer is often preferred by the evolutionary algorithm and allows building cost-efficient networks. The proposed approach discovers new diverse and interesting video architectures that were unknown previously. More importantly they are both more accurate and faster than prior models, and outperform the state-of-the-art results on four datasets: Kinetics, Charades, Moments in Time and HMDB. We will open source the code and models, to encourage future model development.

We present a novel approach for unsupervised learning of depth and ego-motion from monocular video. Unsupervised learning removes the need for separate supervisory signals (depth or ego-motion ground truth, or multi-view video). Prior work in unsupervised depth learning uses pixel-wise or gradient-based losses, which only consider pixels in small local neighborhoods. Our main contribution is to explicitly consider the inferred 3D geometry of the whole scene, and enforce consistency of the estimated 3D point clouds and ego-motion across consecutive frames. This is a challenging task and is solved by a novel (approximate) backpropagation algorithm for aligning 3D structures. We combine this novel 3D-based loss with 2D losses based on photometric quality of frame reconstructions using estimated depth and ego-motion from adjacent frames. We also incorporate validity masks to avoid penalizing areas in which no useful information exists. We test our algorithm on the KITTI dataset and on a video dataset captured on an uncalibrated mobile phone camera. Our proposed approach consistently improves depth estimates on both datasets, and outperforms the state-of-the-art for both depth and ego-motion. Because we only require a simple video, learning depth and ego-motion on large and varied datasets becomes possible. We demonstrate this by training on the low quality uncalibrated video dataset and evaluating on KITTI, ranking among top performing prior methods which are trained on KITTI itself.

We approach structured output prediction by optimizing a deep value network (DVN) to precisely estimate the task loss on different output configurations for a given input. Once the model is trained, we perform inference by gradient descent on the continuous relaxations of the output variables to find outputs with promising scores from the value network. When applied to image segmentation, the value network takes an image and a segmentation mask as inputs and predicts a scalar estimating the intersection over union between the input and ground truth masks. For multi-label classification, the DVN’s objective is to correctly predict the F1 score for any potential label configuration. The DVN framework achieves the state-of-the-art results on multi-label prediction and image segmentation benchmarks.

Learning with deep neural networks forms the state-of-the-art in many tasks such as image classification, image detection, speech recognition, text analysis. We here set out to gain understanding in learning in an `end-to-end' manner for an autonomous vehicle, which refers to directly learning the decision which will result from the perception of the scene. For example, we consider learning a binary `stop'/`go' decision, with respect to pedestrians, given the input image. In this work we propose to use additional information, referred to as `proxy supervision', for improved learning and study its effects on the overall performance. We show that the proxy labels significantly improve the robustness of learning, while achieving as good, or better, accuracy than in the original task of binary classification.

We consider the problem of next frame prediction from video input. A recurrent convolutional neural network is trained to predict depth from monocular video input, which, along with the current video image and the camera trajectory, can then be used to compute the next frame. Unlike prior next-frame prediction approaches, we take advantage of the scene geometry and use the predicted depth for generating the next frame prediction. Our approach can produce rich next frame predictions which include depth information attached to each pixel. Another novel aspect of our approach is that it predicts depth from a sequence of images (e.g. in a video), rather than from a single still image. We evaluate the proposed approach on the KITTI dataset, a standard dataset for benchmarking tasks relevant to autonomous driving. The proposed method produces results which are visually and numerically superior to existing methods that directly predict the next frame. We show that the accuracy of depth prediction improves as more prior frames are considered.

We present a new real-time approach to object detection that exploits the efficiency of cascade classifiers with the accuracy of deep neural networks. Deep networks have been shown to excel at classification tasks, and their ability to operate on raw pixel input without the need to design special features is very appealing. However, deep nets are notoriously slow at inference time. In this paper, we propose an approach that cascades deep nets and fast features, that is both very fast and very accurate. We apply it to the challenging task of pedestrian detection. Our algorithm runs in real-time at 15 frames per second. The resulting approach achieves a 26.2% average miss rate on the Caltech Pedestrian detection benchmark, which is competitive with the very best reported results. It is the first work we are aware of that achieves very high accuracy while running in real-time.

Pedestrian detection is of crucial importance to autonomous driving applications. Methods based on deep learning have shown significant improvements in accuracy, which makes them particularly suitable for applications, such as pedestrian detection, where reducing the miss rate is very important. Although they are accurate, their runtime has been at best in seconds per image, which makes them not practical for onboard applications. We present a Large-Field-Of-View (LFOV) deep network for pedestrian detection, that can achieve high accuracy and is designed to make deep networks work faster for detection problems. The idea of the proposed Large-Field-of-View deep network is to learn to make classification decisions simultaneously and accurately at multiple locations. The LFOV network processes larger image areas at much faster speeds than typical deep networks have been able to, and can intrinsically reuse computations. Our pedestrian detection solution, which is a combination of a LFOV network and a standard deep network, works at 280 ms per image on GPU and achieves 35.85 average miss rate on the Caltech Pedestrian Detection Benchmark.

We present an accurate, real-time approach to robotic grasp detection based on convolutional neural networks. Our network performs single-stage regression to graspable bounding boxes without using standard sliding window or region proposal techniques. The model outperforms state-of-the-art approaches by 14 percentage points and runs at 13 frames per second on a GPU. Our network can simultaneously perform classification so that in a single step it recognizes the object and finds a good grasp rectangle. A modification to this model predicts multiple grasps per object by using a locally constrained prediction mechanism. The locally constrained model performs significantly better, especially on objects that can be grasped in a variety of ways.

This paper presents a systematic evaluation of recent methods in the fine-grained categorization domain, which have shown significant promise. More specifically, we investigate an automatic segmentation algorithm, a region pooling algorithm which is akin to pose-normalized pooling [31] [28], and a multi-class optimization method. We considered the largest and most popular datasets for fine-grained categorization available in the field: the Caltech-UCSD 200 Birds dataset [27], the Oxford 102 Flowers dataset [19], the Stanford 120 Dogs dataset [16], and the Oxford 37 Cats and Dogs dataset [21]. We view this work from a practitioner's perspective, answering the question: what are the methods that can create the best possible fine-grained recognition system which can be applied in practice? Our experiments provide insights of the relative merit of these methods. More importantly, after combining the methods, we achieve the top results in the field, outperforming the state-of-the-art methods by 4.8% and 10.3% for birds and dogs datasets, respectively. Additionally, our method achieves a mAP of 37.92 on the of 2012 Imagenet Fine-Grained Categorization Challenge [1], which outperforms the winner of this challenge by 5.7 points.

This paper addresses the problem of fine-grained recognition in which local, mid-level features are used for classification. We propose to use the Multi-Kernel Learning framework to learn the relative importance of the features and to select optimal features with regards to the classification performance, in a principled way. Our results show improved classification results on common benchmarks for fine-grained classification, as compared to the best prior state-of-the-art methods. The proposed learning-based combination method also improves the concatenation combination approach which has been the standard practice in combining features so far.

This paper addresses a general feature indexing and retrieval scenario in which a set of features detected in the image can retrieve a relevant class of objects, or classes of objects. The main idea behind those features for general object retrieval is that they are capable of identifying and localizing some small regions or parts of the potential object. We propose a set of criteria which take advantage of the learned features to find regions in the image which likely belong to an object. We further use the features' localization capability to localize the full object of interest and its extents. The proposed approach improves the recognition performance and is very efficient. Moreover, it has the potential to be used in automatic image understanding or annotation since it can uncover regions where the objects can be found in an image.

We propose a detection and segmentation algorithm for the purposes of fine-grained recognition. The algorithm first detects low-level regions that could potentially belong to the object and then performs a full-object segmentation through propagation. Apart from segmenting the object, we can also `zoom in' on the object, i.e. center it, normalize it for scale, and thus discount the effects of the background. We then show that combining this with a state-of-the-art classification algorithm leads to significant improvements in performance especially for datasets which are considered particularly hard for recognition, e.g. birds species. The proposed algorithm is much more efficient than other known methods in similar scenarios. Our method is also simpler and we apply it here to different classes of objects, e.g. birds, flowers, cats and dogs. We tested the algorithm on a number of benchmark datasets for fine-grained categorization. It outperforms all the known state-of-the-art methods on these datasets, sometimes by as much as 11%. It improves the performance of our baseline algorithm by 3-4%, consistently on all datasets. We also observed more than a 4% improvement in the recognition performance on a challenging large-scale flower dataset, containing 578 species of flowers and 250,000 images.

We propose a segmentation algorithm for the purposes of large-scale flower species recognition. Our approach is based on identifying potential object regions at the time of detection. We then apply a Laplacian-based segmentation, which is guided by these initially detected regions. More specifically, we show that 1) recognizing parts of the potential object helps the segmentation and makes it more robust to variabilities in both the background and the object appearances, 2) segmenting the object of interest at test time is beneficial for the subsequent recognition. Here we consider a large-scale dataset containing 578 flower species and 250,000 images. This dataset is developed by our team for the purposes of providing a flower recognition application for general use and is the largest in its scale and scope. We tested the proposed segmentation algorithm on the well-known 102 Oxford flowers benchmark [11] and on the new challenging large-scale 578 flower dataset, that we have collected. We observed about 4% improvements in the recognition performance on both datasets compared to the baseline. The algorithm also improves all other known results on the Oxford 102 flower benchmark dataset. Furthermore, our method is both simpler and faster than other related approaches, e.g. [3, 14], and can be potentially applicable to other subcategory recognition datasets.

We propose a method for learning using a set of feature representations which retrieve different amounts of information at different costs. The goal is to create a more efficient terrain classification algorithm which can be used in real-time, onboard an autonomous vehicle. Instead of building a monolithic classifier with uniformly complex representation for each class, the main idea here is to actively consider the labels or misclassification cost while constructing the classifier. For example, some terrain classes might be easily separable from the rest, so very simple representation will be sufficient to learn and detect these classes. This is taken advantage of during learning, so the algorithm automatically builds a variable-length visual representation which varies according to the complexity of the classification task. This enables fast recognition of different terrain types during testing. We also show how to select a set of feature representations so that the desired terrain classification task is accomplished with high accuracy and is at the same time efficient. The proposed approach achieves a good trade-off between recognition performance and speedup on data collected by an autonomous robot.

We address the problem of learning terrain traversability properties from visual input, using automatic mechanical supervision collected from sensors onboard an autonomous vehicle. We present a novel probabilistic framework in which the visual information and the mechanical supervision interact to learn particular terrain types and their properties. The proposed method is applied to learning of rover slippage from visual information in a completely automatic fashion. Our experiments show that using mechanical measurements as automatic supervision significantly improves the visual-based classification alone and approaches the results of learning with manual supervision. This work will enable the rover to drive safely on slopes, learning autonomously about different terrains and their slip characteristics.

In this paper we predict the amount of slip an exploration rover would experience using stereo imagery by learning from previous examples of traversing similar terrain. To do that, the information of terrain appearance and geometry regarding some location is correlated to the slip measured by the rover while this location is being traversed. This relationship is learned from previous experience, so slip can be predicted later at a distance from visual information only. The advantages of the approach are: 1) learning from examples allows the system to adapt to unknown terrains rather than using fixed heuristics or predefined rules; 2) the feedback about the observed slip is received from the vehicle's own sensors which can fully automate the process; 3) learning slip from previous experience can replace complex mechanical modeling of vehicle or terrain, which is time consuming and not necessarily feasible. Predicting slip is motivated by the need to assess the risk of getting trapped before entering a particular terrain. For example, a planning algorithm can utilize slip information by taking into consideration that a slippery terrain is costly or hazardous to traverse. A generic nonlinear regression framework is proposed in which the terrain type is determined from appearance and then a nonlinear model of slip is learned for a particular terrain type. In this paper we focus only on the latter problem and provide slip learning and prediction results for terrain types, such as soil, sand, gravel, and asphalt. The slip prediction error achieved is about 15% which is comparable to the measurement errors for slip itself

Training datasets for learning of object categories are often contaminated or imperfect. We explore an approach to automatically identify examples that are noisy or troublesome for learning and exclude them from the training set. The problem is relevant to learning in semi-supervised or unsupervised setting, as well as to learning when the training data is contaminated with wrongly labeled examples or when correctly labeled, but hard to learn examples, are present. We propose a fully automatic mechanism for noise cleaning, called 'data pruning' and demonstrate its success on learning of human faces. It is not assumed that the data or the noise can be modeled or that additional training examples are available. Our experiments show that data pruning can improve on generalization performance for algorithms with various robustness to noise. It outperforms methods with regularization properties and is superior to commonly applied aggregation methods, such as bagging.