no code implementations • 23 Feb 2024 • Anja Delić, Matej Grcić, Siniša Šegvić
We therefore reconsider direct prediction of K+1 logits that correspond to K groundtruth classes and one outlier class.
1 code implementation • 20 Aug 2023 • Masoud Taghikhah, Nishant Kumar, Siniša Šegvić, Abouzar Eslami, Stefan Gumhold
Previous attempts to address this challenge involved training image classifiers through contrastive learning using actual outlier data or synthesizing outliers for self-supervised learning.
no code implementations • 24 May 2023 • Anja Delić, Matej Grcić, Siniša Šegvić
Most approaches to dense anomaly detection rely on generative modeling or on discriminative methods that train with negative data.
no code implementations • 13 Mar 2023 • Marius Schubert, Tobias Riedlinger, Karsten Kahl, Daniel Kröll, Sebastian Schoenen, Siniša Šegvić, Matthias Rottmann
In this work, we for the first time introduce a benchmark for label error detection methods on object detection datasets as well as a label error detection method and a number of baselines.
1 code implementation • CVPR 2023 • Nishant Kumar, Siniša Šegvić, Abouzar Eslami, Stefan Gumhold
However, this strategy does not guarantee that the synthesized outlier features will have a low likelihood according to the other class-conditional Gaussians.
no code implementations • 19 Jan 2023 • Matej Grcić, Siniša Šegvić
We evaluate our contributions on benchmarks for dense anomaly detection and open-set segmentation.
1 code implementation • 9 Jan 2023 • Matej Grcić, Josip Šarić, Siniša Šegvić
Most dense recognition approaches bring a separate decision in each particular pixel.
1 code implementation • 20 Dec 2022 • Petra Bevandić, Marin Oršić, Ivan Grubišić, Josip Šarić, Siniša Šegvić
For instance, pickups are labeled as trucks in VIPER, cars in Vistas, and vans in ADE20k.
no code implementations • 8 Nov 2022 • Marin Kačan, Marko Ševrović, Siniša Šegvić
We also validate our approach by comparing it with the related work on two road-scene classification datasets from the literature: Honda Scenes and FM3m.
no code implementations • 18 Jul 2022 • Petra Bevandić, Siniša Šegvić
Training semantic segmentation models on multiple datasets has sparked a lot of recent interest in the computer vision community.
1 code implementation • 6 Jul 2022 • Matej Grcić, Petra Bevandić, Siniša Šegvić
We blend these two predictions into a hybrid anomaly score which allows dense open-set recognition on large natural images.
Ranked #3 on Scene Segmentation on StreetHazards (using extra training data)
1 code implementation • 15 Mar 2022 • Josip Šarić, Marin Oršić, Siniša Šegvić
Dense panoptic prediction is a key ingredient in many existing applications such as autonomous driving, automated warehouses or remote sensing.
no code implementations • 23 Dec 2021 • Matej Grcić, Petra Bevandić, Zoran Kalafatić, Siniša Šegvić
Standard machine learning is unable to accommodate inputs which do not belong to the training distribution.
Ranked #2 on Anomaly Detection on Fishyscapes L&F (using extra training data)
1 code implementation • 25 Aug 2021 • Petra Bevandić, Marin Oršić, Ivan Grubišić, Josip Šarić, Siniša Šegvić
Deep supervised models have an unprecedented capacity to absorb large quantities of training data.
1 code implementation • 13 Jun 2021 • Ivan Grubišić, Marin Oršić, Siniša Šegvić
Our experiments show clear advantages of (1) one-way consistency, (2) perturbing only the student branch, and (3) strong photometric and geometric perturbations.
1 code implementation • NeurIPS 2021 • Matej Grcić, Ivan Grubišić, Siniša Šegvić
Normalizing flows are bijective mappings between inputs and latent representations with a fully factorized distribution.
Ranked #2 on Image Generation on ImageNet 32x32 (bpd metric)
no code implementations • 26 Jan 2021 • Josip Šarić, Sacha Vražić, Siniša Šegvić
Feature-to-feature (F2F) module regresses the future features directly and is therefore able to account for emergent scenery.
no code implementations • 22 Jan 2021 • Petra Bevandić, Ivan Krešo, Marin Oršić, Siniša Šegvić
Deep convolutional models often produce inadequate predictions for inputs foreign to the training distribution.
1 code implementation • 22 Nov 2020 • Matej Grcić, Petra Bevandić, Siniša Šegvić
We obtain the synthetic outliers by sampling an RNVP model which is jointly trained to generate datapoints at the border of the training distribution.
no code implementations • 18 Oct 2020 • Kristijan Fugošić, Josip Šarić, Siniša Šegvić
Most existing approaches address this problem as deterministic regression of future features or future predictions given observed frames.
no code implementations • 2 Sep 2020 • Petra Bevandić, Marin Oršić, Ivan Grubišić, Josip Šarić, Siniša Šegvić
We present our submission to the semantic segmentation contest of the Robust Vision Challenge held at ECCV 2020.
1 code implementation • 3 Aug 2019 • Petra Bevandić, Ivan Krešo, Marin Oršić, Siniša Šegvić
Recent success on realistic road driving datasets has increased interest in exploring robust performance in real-world applications.
Ranked #14 on Anomaly Detection on Fishyscapes L&F
no code implementations • 26 Jul 2019 • Josip Šarić, Marin Oršić, Tonći Antunović, Sacha Vražić, Siniša Šegvić
We present a method to anticipate semantic segmentation of future frames in driving scenarios based on feature-to-feature forecasting.
no code implementations • 16 Jul 2019 • Borna Bićanić, Marin Oršić, Ivan Marković, Siniša Šegvić, Ivan Petrović
We investigate tracking-by-detection approaches based on a deep learning detector, joint integrated probabilistic data association (JIPDA), and appearance-based tracking of deep correspondence embeddings.
3 code implementations • 14 May 2019 • Ivan Krešo, Josip Krapac, Siniša Šegvić
Recent progress of deep image classification models has provided great potential to improve state-of-the-art performance in related computer vision tasks.
6 code implementations • 20 Mar 2019 • Marin Oršić, Ivan Krešo, Petra Bevandić, Siniša Šegvić
Recent success of semantic segmentation approaches on demanding road driving datasets has spurred interest in many related application fields.
Ranked #9 on Semantic Segmentation on ZJU-RGB-P
no code implementations • ICLR 2019 • Petra Bevandić, Ivan Krešo, Marin Oršić, Siniša Šegvić
Most classification and segmentation datasets assume a closed-world scenario in which predictions are expressed as distribution over a predetermined set of visual classes.
no code implementations • 9 Jun 2018 • Ivan Krešo, Marin Oršić, Petra Bevandić, Siniša Šegvić
We present semantic segmentation experiments with a model capable to perform predictions on four benchmark datasets: Cityscapes, ScanNet, WildDash and KITTI.
no code implementations • 1 Oct 2013 • Valentina Zadrija, Siniša Šegvić
We consider the problem of multiclass road sign detection using a classification function with multiplicative kernel comprised from two kernels.
no code implementations • 1 Oct 2013 • Ivan Krešo, Marko Ševrović, Siniša Šegvić
In this work, we present a novel dataset for assessing the accuracy of stereo visual odometry.
no code implementations • 1 Oct 2013 • Karla Brkić, Srđan Rašić, Axel Pinz, Siniša Šegvić, Zoran Kalafatić
This paper proposes combining spatio-temporal appearance (STA) descriptors with optical flow for human action recognition.
no code implementations • 1 Oct 2013 • Ivan Sikirić, Karla Brkić, Siniša Šegvić
This paper investigates classification of traffic scenes in a very low bandwidth scenario, where an image should be coded by a small number of features.
no code implementations • 1 Oct 2013 • Sven Lončarić, Siniša Šegvić
Proceedings of the Second Croatian Computer Vision Workshop (CCVW 2013, http://www. fer. unizg. hr/crv/ccvw2013) held September 19, 2013, in Zagreb, Croatia.