11 October 2023 in Multimedia Tools and Applications
5 June 2025 in IEEE Access
A novelty detection task involves identifying whether a data point is an outlier, given a training dataset that primarily captures the distribution of inliers. The novel class is usually absent, poorly sampled, or not well defined in the training data. A common technique for anomaly detection at present is to use an adversarial network generator to generate an anomaly score for inputs using the reconstruction loss. However, because this technique uses a competitive training process, it can be unreliable, with its performance being inconsistent during each adversarial training step. This inconsistency arises from changes in the network’s ability to detect anomalies. In this paper, we propose a revised framework for generative probabilistic novelty detection. We use a similar adversarial autoencoder-based framework but with a lightweight deep network, a novel training paradigm, and a probabilistic score to compute the reconstruction loss. Our methodology calculates the probability of whether a sample comes from the inlier distribution or not. The proposed approach can be applied to anomaly and outlier detection in images and videos. We present the results on multiple benchmark datasets, including the challenging UCSD Ped2 dataset for video anomaly detection. Our results illustrate that our proposed method learns the inlier classes and differentiates them from the outlier classes effectively, leading to better results than the baseline and state-of-the-art methods in several benchmark datasets.
1 January 2026 in IET Computer Vision
We present TDepth, a lightweight hybrid model for monocular depth estimation trained using teacher‐generated pseudo labels. TDepth combines a small CNN backbone, a local‐global transformer encoder, an attention‐guided skip‐fusion decoder, a bin‐based depth head and a lightweight refiner for improving boundaries and thin structures. To avoid dependence on expensive sensor depth during training, we generate pseudo depth labels offline using Pixel‐Perfect‐Depth and train on about 550K images from NYU Depth v2 and ScanNet++. We evaluate TDepth on in‐domain, zero‐shot and efficiency settings. On NYU Depth v2, TDepth‐Full achieves , AbsRel and RMSE , while the smaller 15.1M‐parameter variant remains competitive with much lower model complexity. In zero‐shot evaluation, TDepth‐Full achieves and AbsRel on iBims‐1 and and AbsRel on SUN‐RGBD, showing strong transfer to unseen indoor datasets. TDepth‐Small reaches about 240 FPS at and remains above 30 FPS even at , while TDepth‐Full reaches 60 FPS at and provides stronger depth accuracy with a practical runtime profile. These results show that TDepth provides a strong balance between depth accuracy, zero‐shot generalisation and computational efficiency, making it a practical option for lightweight monocular depth estimation.
27 April 2026 in IEEE Access
Camera rigs for Advanced Driver Assistance Systems (ADAS) are subject to design choices of each vehicle variant; therefore, ADAS algorithms should be rig-invariant or tunable with limited additional data. Recently, the ADAS community adopted BEV (Bird’s Eye View) perception and planning systems due to their scalability to heterogeneous sensors, computational efficiency, and ego-centric perception, but BEV methods remain highly rig-dependent. To verify this, we study a camera-only BEV model, GaussianLSS, which lifts images into a 3D Gaussian scene and predicts BEV maps. We observe a sharp performance drop under a Sedan-to-SUV rig change, with a similar degradation in the reverse direction. Overlap between predicted and ground-truth BEV maps, measured by intersection-over-union (IoU), falls from 38.42% to 18.30% when trained on Sedan and tested on SUV. To address this, we propose a training-time extension to GaussianLSS that improves cross-rig robustness without using any target-rig images or BEV labels. Starting from sedan images, GaussianLSS lifts features into a 3D Gaussian scene, which we render into rig-aware perspective features for both Sedan and SUV. A lightweight decoder is trained only on sedan perspective features to reconstruct sedan RGB (re-projecting features back to the input image) and is then frozen. During cyclic fine-tuning, SUV perspective features rendered from the same Gaussians are passed through the frozen decoder to produce pseudo-SUV views that provide self-supervision for the BEV model in world/BEV space. This increases SUV performance to 23.46% IoU while keeping Sedan performance near its upper bound, and to 25.15% when the decoder additionally sees 20% SUV RGB (the BEV model still uses no SUV labels). The decoder is not used at inference, keeping deployment cost unchanged while suggests a path toward reducing the need for per-vehicle data collection and annotation.