2025
Joint Classification and Unknown Detection Using Class Conditional Probability Calibration
ISSCS 2025
Daniel Brignac, Adam Cuellar, Banafsheh Latibari, Abhijit Mahalanobis
The closed set assumption, where training classes are fixed at inference, is often impractical as deployed models face open-set conditions with unknown classes. This challenge drives the field of Open-Set Recognition (OSR), which aims to identify unknown samples during inference. A common approach to OSR involves training on exemplars of unknown objects (also referred to as known unknowns), which are examples that do not belong to the closed set of known classes. However, this is infeasible for methods that rely solely on the training samples of the known classes. For such cases, we show that the OSR problem can be effectively tackled by combining the decision confidences of two networks: one trained with softmax cross-entropy and the other with tuplet loss using class anchors. We show that the proposed approach outperforms individual methods across OSR benchmarks, maintaining correct classification and high confidence for known samples while effectively rejecting unknowns.
[Link]
Modern novel view synthesis algorithms: a survey
SPIE 2025
Alex Berian, Abhijit Mahalanobis
Novel-view synthesis (NVS) has advanced rapidly, especially with Neural Radiance Fields (NeRF), which map 3D coordinates to colors and densities via volume rendering. While effective, NeRF depends on dense input images and lacks generalization. Recent approaches have improved NeRF by enhancing image quality, reducing training time, and enabling few-shot NVS, which aims to generate novel views from minimal inputs. Key innovations include PixelNeRF’s scene-agnostic synthesis, geometry-aware priors, and the adaptation of diffusion models like DreamFusion using score-distillation-sampling (SDS). Additional strategies, such as multi-view diffusion models and planar scene representations, further advance few-shot NVS. This survey reviews these developments in few-shot NVS.
[Link]
Simultaneous classification of objects and unknown rejection (SCOUR) for infrared target recognition
SPIE 2025
Adam Cuellar, Daniel Brignac, Abhijit Mahalanobis, Wasfy Mikhael
Recognizing targets in infra-red images is an important problem for defense and security applications. A deployed network must not only recognize the known classes, but it must also reject any new or unknown objects without confusing them to be one of the known classes. Our goal is to enhance the ability of existing (or pretrained) classifiers to detect and reject unknown classes. Specifically, we do not alter the training strategy of the main classifier so that its performance on known classes remains unchanged. Instead, we introduce a second network (trained using regression) that uses the decision of the primary classifier to produce a class conditional score that indicates whether an input object is indeed a known object. This is performed in a Bayesian framework where the classification confidence of the primary network is combined with the class-conditional score of the secondary network to accurately separate the unknown objects from the known target classes. Most importantly, our method does not require any examples of OOD imagery to be used for training the second network. For illustrative purposes, we demonstrate the effectiveness of the proposed method using the CIFAR-10 dataset. Ultimately, our goal is to classify known targets in infra-red images while improving the ability to reject unknown classes. Towards this end, we train and test our method on a public domain medium-wave infra-red (MWIR) dataset provided by the US Army for the development of automatic target recognition (ATR) algorithms. The results of this experiment show that the proposed method outperforms other state-of-the-art methods in rejecting the unknown target types while accurately classifying the known ones.
[Link]
CrossModalityDiffusion: Multi-Modal Novel View Synthesis with Unified Intermediate Representation
WACV 2025
Alex Berian, Daniel Brignac, JhihYang Wu, Natnael Daba, Abhijit Mahalanobis
Geospatial imaging leverages data from diverse sensing modalities-such as EO, SAR, and LiDAR, ranging from ground-level drones to satellite views. These heterogeneous inputs offer significant opportunities for scene understanding but present challenges in interpreting geometry accurately, particularly in the absence of precise ground truth data. To address this, we propose CrossModalityDiffusion, a modular framework designed to generate images across different modalities and viewpoints without prior knowledge of scene geometry. CrossModalityDiffusion employs modality-specific encoders that take multiple input images and produce geometry-aware feature volumes that encode scene structure relative to their input camera positions. The space where the feature volumes are placed acts as a common ground for unifying input modalities. These feature volumes are overlapped and rendered into feature images from novel perspectives using volumetric rendering techniques. The rendered feature images are used as conditioning inputs for a modality-specific diffusion model, enabling the synthesis of novel images for the desired output modality. In this paper, we show that jointly training different modules ensures consistent geometric understanding across all modalities within the framework. We validate CrossModalityDiffusion’s capabilities on the synthetic ShapeNet cars dataset, demonstrating its effectiveness in generating accurate and consistent novel views across multiple imaging modalities and perspectives.
[PDF] [Code]
2024
Cascading Unknown Detection with Known Classification for Open Set Recognition
ICIP 2024
Daniel Brignac, Abhijit Mahalanobis
Deep learners tend to perform well when trained under the closed set assumption but struggle when deployed under open set conditions. This motivates the field of Open Set Recognition (OSR) in which we seek to give deep learners the ability to recognize whether a data sample belongs to the known classes trained on or comes from the surrounding infinite world. In this work, we decompose the traditional OSR formulation into fine class distinction and known/unknown class discrimination.
[PDF]
Estimation of Class Priors for Improving Classification Accuracy during Deployment
Advances in Machine Learning and Image Analysis for GeoAI
Bruce McIntosh, Natnael Daba, Abhijit Mahalanobis
Conventional classifiers are trained and evaluated using “balanced” data sets, in which all classes are equally present. However, it is unlikely that the prior probability of occurrence will be the same for the classes that are encountered in each deployment scenario. Unfortunately, the exact class priors are generally unknown and can vary spatially and over time. We explore different methods for estimating the class priors based on the response of the classifier itself. We observe that the distribution of classifier’s decisions made in response to a collection test image can be expressed as a linear combination of the distribution of its decisions made in response to each individual class. The weights of the linear combination are the prior probabilities with which the different classes occur in the test set. This can be used along with the classifier’s precision and recall for each class to estimate the class priors. We then show that incorporating the estimated class priors in the overall decision scheme improves the classifier’s accuracy in the context of the deployment. We first illustrate the potential improvements in classification accuracy using the CIFAR-100 and Tiny ImageNet data sets and then demonstrate the application of this technique for improving the classification and recognition of objects in satellite imagery using the well-known Xview dataset.
[PDF]
2023
Improving Replay Sample Selection and Storage for Less Forgetting in Continual Learning
ICCV 2023
Daniel Brignac, Niels Lobo, Abhijit Mahalanobis
Continual Learners commonly employ replay as a method for mitigating catastrophic forgetting where we store few previous examples in a memory buffer and replay them when learning a new task. Replay commonly uses random sampling strategies (e.g., reservoir sampling) to populate the buffer which can potentially store uninformative/redundant data and overwrite informative data. This works explores replacements to reservoir sampling for less forgetting when using replay methods.
[PDF]
Adapting Classifiers To Changing Class Priors During Deployment
Natnael Daba, Bruce McIntosh, Abhijit Mahalanobis
Conventional classifiers are trained and evaluated using balanced data sets in which all classes are equally present. Classifiers are now trained on large data sets such as ImageNet, and are now able to classify hundreds (if not thousands) of different classes. On one hand, it is desirable to train such general-purpose classifier on a very large number of classes so that it performs well regardless of the settings in which it is deployed. On the other hand, it is unlikely that all classes known to the classifier will occur in every deployment scenario, or that they will occur with the same prior probability. In reality, only a relatively small subset of the known classes may be present in a particular setting or environment. For example, a classifier will encounter mostly animals if its deployed in a zoo or for monitoring wildlife, aircraft and service vehicles at an airport, or various types of automobiles and commercial vehicles if it is used for monitoring traffic. Furthermore, the exact class priors are generally unknown and can vary over time. In this paper, we explore different methods for estimating the class priors based on the output of the classifier itself. We then show that incorporating the estimated class priors in the overall decision scheme enables the classifier to increase its run-time accuracy in the context of its deployment scenario.
[PDF]