This paper investigates the problem of class-incremental object detection for agricultural applications where a model needs to learn new plant species and diseases incrementally without forgetting the previously learned ones. We adapt two public datasets to include new categories over time, simulating a more realistic and dynamic scenario. We then compare three class-incremental learning methods that leverage different forms of knowledge distillation to mitigate catastrophic forgetting. Our experiments show that all three methods suffer from catastrophic forgetting, but the Dynamic Y-KD approach, which additionally uses a dynamic architecture that grows new branches to learn new tasks, outperforms ILOD and Faster-ILOD in most settings both on new and old classes. These results highlight the challenges and opportunities of continual object detection for agricultural applications. In particular, we hypothesize that the large intra-class and small inter-class variability that is typical of plant images exacerbate the difficulty of learning new categories without interfering with previous knowledge. We publicly release our code to encourage future work.
Instance segmentation is a fundamental computer vision problem which remains challenging despite impressive recent advances due to deep learning-based methods. Given sufficient training data, fully supervised methods can yield excellent performance, but annotation of groundtruth remains a major bottleneck, especially for biomedical applications where it has to be performed by domain experts. The amount of labels required can be drastically reduced by using rules derived from prior knowledge to guide the segmentation. However, these rules are in general not differentiable and thus cannot be used with existing methods. Here, we revoke this requirement by using stateless actor critic reinforcement learning, which enables non-differentiable rewards. We formulate the instance segmentation problem as graph partitioning and the actor critic predicts the edge weights driven by the rewards, which are based on the conformity of segmented instances to high-level priors on object shape, position or size. The experiments on toy and real data demonstrate that a good set of priors is sufficient to reach excellent performance without any direct object-level supervision.
Recent works show that the data distribution in a network's latent space is useful for estimating classification uncertainty and detecting Out-of-distribution (OOD) samples. To obtain a well-regularized latent space that is conducive for uncertainty estimation, existing methods bring in significant changes to model architectures and training procedures. In this paper, we present a lightweight, fast, and high-performance regularization method for Mahalanobis distance-based uncertainty prediction, and that requires minimal changes to the network's architecture. To derive Gaussian latent representation favourable for Mahalanobis Distance calculation, we introduce a self-supervised representation learning method that separates in-class representations into multiple Gaussians. Classes with non-Gaussian representations are automatically identified and dynamically clustered into multiple new classes that are approximately Gaussian. Evaluation on standard OOD benchmarks shows that our method achieves state-of-the-art results on OOD detection with minimal inference time, and is very competitive on predictive probability calibration. Finally, we show the applicability of our method to a real-life computer vision use case on microorganism classification.
Lung cancer is a grave disease that accounts for more than one million deaths, and Non-Small Cell Lung Cancer (NSCLC) accounts for 85% of all lung cancers. Rapid detection of lung cancer could reduce the mortality rate and increase the patient's survival rate, in which tumor segmentation plays a significant role in the diagnosis and treatment of lung cancer. Nevertheless, manual segmentation by radiologists can be time-consuming and labor-intensive. In recent years, deep learning methods have achieved good results in medical image segmentation. In this paper, RRcUNet 3D, a variant of the Unet model, was proposed to perform tumor segmentation in Computed Tomography (CT) images of NSCLC patients. This network was trained end-to-end from a small set of CT scans of NSCLC patients, then the trained model was validated on another set of CT scans of NSCLC patients. The experimental results showed that our model can provide a highly accurate segmentation of tumors in the 3D volume of CT images.
Online violence against children has increased globally recently, demanding urgent attention. Competent authorities manually analyze abuse complaints to comprehend crime dynamics and identify patterns. However, the manual analysis of these complaints presents a challenge because it exposes analysts to harmful content during the review process. Given these challenges, we present a novel solution, an automated tool designed to analyze children's sexual abuse reports comprehensively. By automating the analysis process, our tool significantly reduces the risk of exposure to harmful content by categorizing the reports on three dimensions: Subject, Degree of Criminality, and Damage. Furthermore, leveraging our multidisciplinary team's expertise, we introduce a novel approach to annotate the collected data, enabling a more in-depth analysis of the reports. This approach improves the comprehension of fundamental patterns and trends, enabling law enforcement agencies and policymakers to create focused strategies in the fight against children's violence.
The input tokens to Vision Transformers carry little semantic meaning as they are defined as regular equal-sized patches of the input image, regardless of its content. However, processing uniform background areas of an image should not necessitate as much compute as dense, cluttered areas. To address this issue, we propose a dynamic mixed-scale tokenization scheme for ViT, MSViT. Our method introduces a conditional gating mechanism that selects the optimal token scale for every image region, such that the number of tokens is dynamically determined per input. In addition, to enhance the conditional behavior of the gate during training, we introduce a novel generalization of the batch-shaping loss. We show that our gating module is able to learn meaningful semantics despite operating locally at the coarse patch-level. The proposed gating module is lightweight, agnostic to the choice of transformer backbone, and trained within a few epochs with little training overhead. Furthermore, in contrast to token pruning, MSViT does not lose information about the input, thus can be readily applied for dense tasks. We validate MSViT on the tasks of classification and segmentation where it leads to improved accuracy-complexity trade-off.
Handling entirely unknown data is a challenge for any deployed classifier. Classification models are typically trained on a static pre-defined dataset and are kept in the dark for the open unassigned feature space. As a result, they struggle to deal with out-of-distribution data during inference. Addressing this task on the class-level is termed open-set recognition (OSR). However, most OSR methods are inherently limited, as they train closed-set classifiers and only adapt the downstream predictions to OSR. This work presents LORD, a framework to Leverage Open-set Recognition by exploiting unknown Data. LORD explicitly models open space during classifier training and provides a systematic evaluation for such approaches. We identify three model-agnostic training strategies that exploit background data and applied them to well-established classifiers. Due to LORD's extensive evaluation protocol, we consistently demonstrate improved recognition of unknown data. The benchmarks facilitate in-depth analysis across various requirement levels. To mitigate dependency on extensive and costly background datasets, we explore mixup as an off-the-shelf data generation technique. Our experiments highlight mixup's effectiveness as a substitute for background datasets. Lightweight constraints on mixup synthesis further improve OSR performance.
Multi-label classification is an essential task utilized in a wide variety of real-world applications. Multi-label zero-shot learning is a method for classifying images into multiple unseen categories for which no training data is available, while in general zero-shot situations, the test set may include observed classes. The CLIP-Decoder is a novel method based on the state-of-the-art ML-Decoder attention-based head. We introduce multi-modal representation learning in CLIP-Decoder, utilizing the text encoder to extract text features and the image encoder for image feature extraction. Furthermore, we minimize semantic mismatch by aligning image and word embeddings in the same dimension and comparing their respective representations using a combined loss, which comprises classification loss and CLIP loss. This strategy outperforms other methods and we achieve cutting-edge results on zero-shot multilabel classification tasks using CLIP-Decoder. Our method achieves an absolute increase of 3.9% in performance compared to existing methods for zero-shot learning multi-label classification tasks. Additionally, in the generalized zero-shot learning multi-label classification task, our method shows an impressive increase of almost 2.3%.
We propose a general way to integrate procedural knowledge of a domain into deep learning models. We apply it to the case of video prediction, building on top of object-centric deep models and show that this leads to a better performance than using data-driven models alone. We develop an architecture that facilitates latent space disentanglement in order to use the integrated procedural knowledge, and establish a setup that allows the model to learn the procedural interface in the latent space using the downstream task of video prediction. We contrast the performance to a state-of-the-art data-driven approach and show that problems where purely data-driven approaches struggle can be handled by using knowledge about the domain, providing an alternative to simply collecting more data.
Autonomous driving perception techniques are typically based on supervised machine learning models that are trained on real-world street data. A typical training process involves capturing images with a single car model and windshield configuration. However, deploying these trained models on different car types can lead to a domain shift, which can potentially hurt the neural networks performance and violate working ADAS requirements. To address this issue, this paper investigates the domain shift problem further by evaluating the sensitivity of two perception models to different windshield configurations. This is done by evaluating the dependencies between neural network benchmark metrics and optical merit functions by applying a Fourier optics based threat model. Our results show that there is a performance gap introduced by windshields and existing optical metrics used for posing requirements might not be sufficient.
This paper introduces a novel benchmark to study the impact and relationship of built environment elements on pedestrian collision prediction, intending to enhance environmental awareness in autonomous driving systems to prevent pedestrian injuries actively. We introduce a built environment detection task in large-scale panoramic images and a detection-based pedestrian collision frequency prediction task. We propose a baseline method that incorporates a collision prediction module into a state-of-the-art detection model to tackle both tasks simultaneously. Our experiments demonstrate a significant correlation between object detection of built environment elements and pedestrian collision frequency prediction. Our results are a stepping stone towards understanding the interdependencies between built environment conditions and pedestrian safety.
Online Object Detection (OOD) algorithms play a crucial role in dynamic and real-world computer vision applications. In these scenarios, models are trained on a data stream where old class samples are revisited, a phenomenon known as Natural Replay (NR). During training, NR occurs unevenly across object categories, leading to evaluation metrics biased towards the most frequently revisited classes. Existing benchmarks lack proper quantification of NR and depict short-term training scenarios on a single domain. As a result, evaluating generalization capabilities and forgetting rates of models become challenging in OOD. In this paper, we address the challenges surrounding the evaluation of OOD models by proposing two key contributions. Firstly, we define a metric to quantify NR in an OOD scenario and show how NR is related to class specific forgetting. Secondly, we introduce a novel benchmark, EgOAK, which introduces a long-term training scenario that involves frequent domain shifts. It allows the evaluation of models’ generalization capabilities and forgetting of knowledge on past domains. Our results in this OOD setting reveal that Experience Replay, a memory-based method, is particularly effective for better generalization to new domains and for preserving past knowledge. Leveraging replay from memory helps to address the low natural replay rate for rarely revisited classes, resulting in improved adaptability and reliability of models in dynamic environments.
We present a novel real-time capable learning method that jointly perceives a 3D scene's geometry structure and semantic labels. Recent approaches to real-time 3D scene reconstruction mostly adopt a volumetric scheme, where a Truncated Signed Distance Function (TSDF) is directly regressed. However, these volumetric approaches tend to focus on the global coherence of their reconstructions, which leads to a lack of local geometric detail. To overcome this issue, we propose to leverage the latent geometric prior knowledge in 2D image features by explicit depth prediction and anchored feature generation, to refine the occupancy learning in TSDF volume. Besides, we find that this cross-dimensional feature refinement methodology can also be adopted for the semantic segmentation task by utilizing semantic priors. Hence, we proposed an end-to-end cross-dimensional refinement neural network (CDRNet) to extract both 3D mesh and 3D semantic labeling in real time. The experiment results show that this method achieves a state-of-the-art 3D perception efficiency on multiple datasets, which indicates the great potential of our method for industrial applications.
Optimal plant reconstruction is an essential element in automating our future agriculture. Computerized inspection of proper growth, nutrition, or pest infestation has become mandatory in fully autonomous in-door or micro-farm settings, shifting from fixed to moving camera systems. In industrial environments, plant scanning must work efficiently with a limited number of significant images to become economically viable. We present an adaptive learning algorithm for agricultural plant inspection robots, in particular, a specific type of reinforcement learning that we developed for our micro-farming platform created within the EU project ROMI. We suggest a new approach to 3D plant reconstruction by integrating the space carving technique with categorical Deep Q-Networks. Space carving leverages images captured from various positions to create a binary voxel grid, representing the occupied and unoccupied spaces of the scanned object. The proposed method incorporates partial 3D reconstructions of plants obtained through space carving, which get compared to a ground truth model to calculate the reward and guide scanning policies. We explain the algorithmic details and the 3D reconstruction technique in design, implementation, and evaluation. Experimental results confirm our approach’s effectiveness in improving the 3D plant reconstruction process, highlighting its potential for further applications in agriculture and related fields.
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, provide 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 the 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 serve as a benchmark for 3D pose estimation.
Designing metrics for evaluating instance segmentation revolves around comprehensively considering object detection and segmentation accuracy. However, other important properties, such as sensitivity, continuity, and equality, are overlooked in the current study. In this paper, we reveal that most existing metrics have a limited resolution of segmentation quality. They are only conditionally sensitive to the change of masks or false predictions. For certain metrics, the score can change drastically in a narrow range which could provide a misleading indication of the quality gap between results. Therefore, we propose a new metric called sortedAP, which strictly decreases with both object- and pixel-level imperfections and has an uninterrupted penalization scale over the entire domain. We provide the evaluation toolkit and experiment code at https://www.github.com/looooongChen/sortedAP.
Large text-to-image diffusion models have achieved remarkable success in generating diverse, high-quality images. Additionally, these models have been successfully leveraged to edit input images by just changing the text prompt. But when these models are applied to videos, the main challenge is to ensure temporal consistency and coherence across frames. In this paper, we propose InFusion, a framework for zero-shot text-based video editing leveraging large pre-trained image diffusion models. Our framework specifically supports editing of multiple concepts with pixel-level control over diverse concepts mentioned in the editing prompt. Specifically, we inject the difference in features obtained with source and edit prompts from U-Net residual blocks of decoder layers. When these are combined with injected attention features, it becomes feasible to query the source contents and scale edited concepts along with the injection of unedited parts. The editing is further controlled in a fine-grained manner with mask extraction and attention fusion, which cut the edited part from the source and paste it into the denoising pipeline for the editing prompt. Our framework is a low-cost alternative to one-shot tuned models for editing since it does not require training. We demonstrated complex concept editing with a generalised image model (Stable Diffusion v1.5) using LoRA. Adaptation is compatible with all the existing image diffusion techniques. Extensive experimental results demonstrate the effectiveness of existing methods in rendering high-quality and temporally consistent videos.
Editing a local region or a specific object in a 3D scene represented by a NeRF or consistently blending a new realistic object into the scene is challenging, mainly due to the implicit nature of the scene representation. We present Blended-NeRF, a robust and flexible framework for editing a specific region of interest in an existing NeRF scene, based on text prompts, along with a 3D ROI box. Our method leverages a pretrained language-image model to steer the synthesis towards a user-provided text prompt, along with a 3D MLP model initialized on an existing NeRF scene to generate the object and blend it into a specified region in the original scene. We allow local editing by localizing a 3D ROI box in the input scene, and blend the content synthesized inside the ROI with the existing scene using a novel volumetric blending technique. To obtain natural looking and view-consistent results, we leverage existing and new geometric priors and 3D augmentations for improving the visual fidelity of the final result. We test our framework both qualitatively and quantitatively on a variety of real 3D scenes and text prompts, demonstrating realistic multi-view consistent results with much flexibility and diversity compared to the baselines. Finally, we show the applicability of our framework for several 3D editing applications, including adding new objects to a scene, removing/replacing/altering existing objects, and texture conversion.
3D dynamic meshes offer significant potential in various applications, but their usage is still limited by their large file size. We present a novel method that can compress 3D human dynamic meshes effectively by using embedded deformation to extract the underlying transformations of consecutive frames. We target 3D dynamic meshes with changing connectivity which are more versatile compared to traditional mesh animation but also more challenging. To further reduce the transmission size, we propose a novel optimization-based technique to determine a sparse set of key nodes capable of transmitting the transformations efficiently.
Examination of chest X-ray images is currently one of the most important methods for the screening and diagnosis of thoracic diseases and, in some cases, for assessing response to treatment. However, this task is time-consuming and expensive as it requires a detailed visual inspection and interpretation by a trained clinician. In the past decade, several machine learning (ML) methods have been developed to remedy this issue as clinical decision support methods. However, most of these algorithms face challenges like computational feasibility, reliability, and interoperability.In this paper, we develop a unique feature extraction method for chest X-rays by applying the latest topological data analysis (TDA) methods. We observe that normal and abnormal images produce very distinct topological patterns for pneumonia and tuberculosis. By using cubical persistence, we capture these patterns and convert them into powerful feature vectors. By combining with standard ML methods, we obtain a computationally feasible and interpretable model. In our extensive experiments, our model Topo-CXR outperforms state-of-the-art deep learning (DL) models in several benchmark datasets. Unlike most DL models, our proposed Topo-CXR model does not need any data augmentation or pre-processing steps and works perfectly on small datasets. Furthermore, our topological feature vectors can be easily integrated with any future ML and DL models to boost their performance and improve robustness.
