Contrastive Language-Image Pre-training (CLIP) plays an essential role in extracting valuable content information from images across diverse tasks. It aligns textual and visual modalities to comprehend the entire image, including all the details, even those irrelevant to specific tasks. However, for a finer understanding and controlled editing of images, it becomes crucial to focus on specific regions of interest, which can be indicated as points, masks, or boxes by humans or perception models. To fulfill the requirements, we introduce Alpha-CLIP, an enhanced version of CLIP with an auxiliary alpha channel to suggest attentive regions and fine-tuned with constructed millions of RGBA region-text pairs. Alpha-CLIP not only preserves the visual recognition ability of CLIP but also enables precise control over the emphasis of image contents. It demonstrates effectiveness in various tasks, including but not limited to open-world recognition, multimodal large language models, and conditional 2D / 3D generation. It has a strong potential to serve as a versatile tool for image-related tasks.
Instance detection (InsDet) is a long-lasting problem in robotics and computer vision, aiming to detect object instances (predefined by some visual examples) in a cluttered scene. Despite its practical significance, its advancement is overshadowed by Object Detection, which aims to detect objects belonging to some predefined classes. One major reason is that current InsDet datasets are too small in scale by today's standards. For example, the popular InsDet dataset GMU (published in 2016) has only 23 instances, far less than COCO (80 classes), a well-known object detection dataset published in 2014. We are motivated to introduce a new InsDet dataset and protocol. First, we define a realistic setup for InsDet: training data consists of multi-view instance captures, along with diverse scene images allowing synthesizing training images by pasting instance images on them with free box annotations. Second, we release a real-world database, which contains multi-view capture of 100 object instances, and high-resolution (6k x 8k) testing images. Third, we extensively study baseline methods for InsDet on our dataset, analyze their performance and suggest future work. Somewhat surprisingly, using the off-the-shelf class-agnostic segmentation model (Segment Anything Model, SAM) and the self-supervised feature representation DINOv2 performs the best, achieving >10 AP better than end-to-end trained InsDet models that repurpose object detectors (e.g., FasterRCNN and RetinaNet).
Trained on web-scale image-text pairs, Vision-Language Models (VLMs) such as CLIP can recognize images of common objects in a zero-shot fashion. However, it is underexplored how to use CLIP for zero-shot recognition of highly specialized concepts, e.g., species of birds, plants, and animals, for which their scientific names are written in Latin or Greek. Indeed, CLIP performs poorly for zero-shot species recognition with prompts that use scientific names, e.g., "a photo of Lepus Timidus" (which is a scientific name in Latin). Because these names are usually not included in CLIP's training set. To improve performance, prior works propose to use large-language models (LLMs) to generate descriptions (e.g., of species color and shape) and additionally use them in prompts. We find that they bring only marginal gains. Differently, we are motivated to translate scientific names (e.g., Lepus Timidus) to common English names (e.g., mountain hare) and use such in the prompts. We find that common names are more likely to be included in CLIP's training set, and prompting them achieves 2$\sim$5 times higher accuracy on benchmarking datasets of fine-grained species recognition.
Segmenting and recognizing diverse object parts is a crucial ability in applications spanning various computer vision and robotic tasks. While significant progress has been made in object-level Open-Vocabulary Semantic Segmentation (OVSS), i.e., segmenting objects with arbitrary text, the corresponding part-level research poses additional challenges. Firstly, part segmentation inherently involves intricate boundaries, while limited annotated data compounds the challenge. Secondly, part segmentation introduces an open granularity challenge due to the diverse and often ambiguous definitions of parts in the open world. Furthermore, the large-scale vision and language models, which play a key role in the open vocabulary setting, struggle to recognize parts as effectively as objects. To comprehensively investigate and tackle these challenges, we propose an Open-Vocabulary Part Segmentation (OV-PARTS) benchmark. OV-PARTS includes refined versions of two publicly available datasets: Pascal-Part-116 and ADE20K-Part-234. And it covers three specific tasks: Generalized Zero-Shot Part Segmentation, Cross-Dataset Part Segmentation, and Few-Shot Part Segmentation, providing insights into analogical reasoning, open granularity and few-shot adapting abilities of models. Moreover, we analyze and adapt two prevailing paradigms of existing object-level OVSS methods for OV-PARTS. Extensive experimental analysis is conducted to inspire future research in leveraging foundational models for OV-PARTS. The code and dataset are available at https://github.com/OpenRobotLab/OV_PARTS.
Knowledge distillation (KD) exploits a large well-trained model (i.e., teacher) to train a small student model on the same dataset for the same task. Treating teacher features as knowledge, prevailing methods of knowledge distillation train student by aligning its features with the teacher's, e.g., by minimizing the KL-divergence between their logits or L2 distance between their intermediate features. While it is natural to believe that better alignment of student features to the teacher better distills teacher knowledge, simply forcing this alignment does not directly contribute to the student's performance, e.g., classification accuracy. In this work, we propose to align student features with class-mean of teacher features, where class-mean naturally serves as a strong classifier. To this end, we explore baseline techniques such as adopting the cosine distance based loss to encourage the similarity between student features and their corresponding class-means of the teacher. Moreover, we train the student to produce large-norm features, inspired by other lines of work (e.g., model pruning and domain adaptation), which find the large-norm features to be more significant. Finally, we propose a rather simple loss term (dubbed ND loss) to simultaneously (1) encourage student to produce large-\emph{norm} features, and (2) align the \emph{direction} of student features and teacher class-means. Experiments on standard benchmarks demonstrate that our explored techniques help existing KD methods achieve better performance, i.e., higher classification accuracy on ImageNet and CIFAR100 datasets, and higher detection precision on COCO dataset. Importantly, our proposed ND loss helps the most, leading to the state-of-the-art performance on these benchmarks. The source code is available at \url{https://github.com/WangYZ1608/Knowledge-Distillation-via-ND}.
We focus on the task of far-field 3D detection (Far3Det) of objects beyond a certain distance from an observer, e.g., $>$50m. Far3Det is particularly important for autonomous vehicles (AVs) operating at highway speeds, which require detections of far-field obstacles to ensure sufficient braking distances. However, contemporary AV benchmarks such as nuScenes underemphasize this problem because they evaluate performance only up to a certain distance (50m). One reason is that obtaining far-field 3D annotations is difficult, particularly for lidar sensors that produce very few point returns for far-away objects. Indeed, we find that almost 50% of far-field objects (beyond 50m) contain zero lidar points. Secondly, current metrics for 3D detection employ a "one-size-fits-all" philosophy, using the same tolerance thresholds for near and far objects, inconsistent with tolerances for both human vision and stereo disparities. Both factors lead to an incomplete analysis of the Far3Det task. For example, while conventional wisdom tells us that high-resolution RGB sensors should be vital for 3D detection of far-away objects, lidar-based methods still rank higher compared to RGB counterparts on the current benchmark leaderboards. As a first step towards a Far3Det benchmark, we develop a method to find well-annotated scenes from the nuScenes dataset and derive a well-annotated far-field validation set. We also propose a Far3Det evaluation protocol and explore various 3D detection methods for Far3Det. Our result convincingly justifies the long-held conventional wisdom that high-resolution RGB improves 3D detection in the far-field. We further propose a simple yet effective method that fuses detections from RGB and lidar detectors based on non-maximum suppression, which remarkably outperforms state-of-the-art 3D detectors in the far-field.
Contemporary autonomous vehicle (AV) benchmarks have advanced techniques for training 3D detectors, particularly on large-scale lidar data. Surprisingly, although semantic class labels naturally follow a long-tailed distribution, contemporary benchmarks focus on only a few common classes (e.g., pedestrian and car) and neglect many rare classes in-the-tail (e.g., debris and stroller). However, AVs must still detect rare classes to ensure safe operation. Moreover, semantic classes are often organized within a hierarchy, e.g., tail classes such as child and construction-worker are arguably subclasses of pedestrian. However, such hierarchical relationships are often ignored, which may lead to misleading estimates of performance and missed opportunities for algorithmic innovation. We address these challenges by formally studying the problem of Long-Tailed 3D Detection (LT3D), which evaluates on all classes, including those in-the-tail. We evaluate and innovate upon popular 3D detection codebases, such as CenterPoint and PointPillars, adapting them for LT3D. We develop hierarchical losses that promote feature sharing across common-vs-rare classes, as well as improved detection metrics that award partial credit to "reasonable" mistakes respecting the hierarchy (e.g., mistaking a child for an adult). Finally, we point out that fine-grained tail class accuracy is particularly improved via multimodal fusion of RGB images with LiDAR; simply put, small fine-grained classes are challenging to identify from sparse (lidar) geometry alone, suggesting that multimodal cues are crucial to long-tailed 3D detection. Our modifications improve accuracy by 5% AP on average for all classes, and dramatically improve AP for rare classes (e.g., stroller AP improves from 3.6 to 31.6)!
Lifelong learners must recognize concept vocabularies that evolve over time. A common yet underexplored scenario is learning with class labels over time that refine/expand old classes. For example, humans learn to recognize ${\tt dog}$ before dog breeds. In practical settings, dataset $\textit{versioning}$ often introduces refinement to ontologies, such as autonomous vehicle benchmarks that refine a previous ${\tt vehicle}$ class into ${\tt school-bus}$ as autonomous operations expand to new cities. This paper formalizes a protocol for studying the problem of $\textit{Learning with Evolving Class Ontology}$ (LECO). LECO requires learning classifiers in distinct time periods (TPs); each TP introduces a new ontology of "fine" labels that refines old ontologies of "coarse" labels (e.g., dog breeds that refine the previous ${\tt dog}$). LECO explores such questions as whether to annotate new data or relabel the old, how to leverage coarse labels, and whether to finetune the previous TP's model or train from scratch. To answer these questions, we leverage insights from related problems such as class-incremental learning. We validate them under the LECO protocol through the lens of image classification (CIFAR and iNaturalist) and semantic segmentation (Mapillary). Our experiments lead to surprising conclusions; while the current status quo is to relabel existing datasets with new ontologies (such as COCO-to-LVIS or Mapillary1.2-to-2.0), LECO demonstrates that a far better strategy is to annotate $\textit{new}$ data with the new ontology. However, this produces an aggregate dataset with inconsistent old-vs-new labels, complicating learning. To address this challenge, we adopt methods from semi-supervised and partial-label learning. Such strategies can surprisingly be made near-optimal, approaching an "oracle" that learns on the aggregate dataset exhaustively labeled with the newest ontology.
Shoe tread impressions are one of the most common types of evidence left at crime scenes. However, the utility of such evidence is limited by the lack of databases of footwear impression patterns that cover the huge and growing number of distinct shoe models. We propose to address this gap by leveraging shoe tread photographs collected by online retailers. The core challenge is to predict the impression pattern from the shoe photograph since ground-truth impressions or 3D shapes of tread patterns are not available. We develop a model that performs intrinsic image decomposition (predicting depth, normal, albedo, and lighting) from a single tread photo. Our approach, which we term ShoeRinsics, combines domain adaptation and re-rendering losses in order to leverage a mix of fully supervised synthetic data and unsupervised retail image data. To validate model performance, we also collected a set of paired shoe-sole images and corresponding prints, and define a benchmarking protocol to quantify the accuracy of predicted impressions. On this benchmark, ShoeRinsics outperforms existing methods for depth prediction and synthetic-to-real domain adaptation.
In the real open world, data tends to follow long-tailed class distributions, motivating the well-studied long-tailed recognition (LTR) problem. Naive training produces models that are biased toward common classes in terms of higher accuracy. The key to addressing LTR is to balance various aspects including data distribution, training losses, and gradients in learning. We explore an orthogonal direction, weight balancing, motivated by the empirical observation that the naively trained classifier has "artificially" larger weights in norm for common classes (because there exists abundant data to train them, unlike the rare classes). We investigate three techniques to balance weights, L2-normalization, weight decay, and MaxNorm. We first point out that L2-normalization "perfectly" balances per-class weights to be unit norm, but such a hard constraint might prevent classes from learning better classifiers. In contrast, weight decay penalizes larger weights more heavily and so learns small balanced weights; the MaxNorm constraint encourages growing small weights within a norm ball but caps all the weights by the radius. Our extensive study shows that both help learn balanced weights and greatly improve the LTR accuracy. Surprisingly, weight decay, although underexplored in LTR, significantly improves over prior work. Therefore, we adopt a two-stage training paradigm and propose a simple approach to LTR: (1) learning features using the cross-entropy loss by tuning weight decay, and (2) learning classifiers using class-balanced loss by tuning weight decay and MaxNorm. Our approach achieves the state-of-the-art accuracy on five standard benchmarks, serving as a future baseline for long-tailed recognition.