1. |
- Moberg, John, et al.
(author)
-
Bayesian Linear Regression on Deep Representations
- 2019
-
In: Advances in Neural Information Processing Systems 32.
-
Conference paper (peer-reviewed)abstract
- A simple approach to obtaining uncertainty-aware neural networks for regression is to do Bayesian linear regression (BLR) on the representation from the last hidden layer. Recent work [Riquelme et al., 2018, Azizzadenesheli et al., 2018] indicates that the method is promising, though it has been limited to homoscedastic noise. In this paper, we propose a novel variation that enables the method to flexibly model heteroscedastic noise. The method is benchmarked against two prominent alternative methods on a set of standard datasets, and finally evaluated as an uncertainty-aware model in model-based reinforcement learning. Our experiments indicate that the method is competitive with standard ensembling, and ensembles of BLR outperforms the methods we compared to.
|
|
2. |
- Olsson, Viktor, et al.
(author)
-
ClassMix: Segmentation-Based Data Augmentation for Semi-Supervised Learning
- 2021
-
In: Proceedings - 2021 IEEE Winter Conference on Applications of Computer Vision, WACV 2021.
-
Conference paper (peer-reviewed)abstract
- The state of the art in semantic segmentation is steadily increasing in performance, resulting in more precise and reliable segmentations in many different applications. However, progress is limited by the cost of generating labels for training, which sometimes requires hours of manual labor for a single image. Because of this, semi-supervised methods have been applied to this task, with varying degrees of success. A key challenge is that common augmentations used in semi-supervised classification are less effective for semantic segmentation. We propose a novel data augmentation mechanism called ClassMix, which generates augmentations by mixing unlabelled samples, by leveraging on the network's predictions for respecting object boundaries. We evaluate this augmentation technique on two common semi-supervised semantic segmentation benchmarks, showing that it attains state-of-the-art results. Lastly, we also provide extensive ablation studies comparing different design decisions and training regimes.
|
|
3. |
- Pinto, Juliano, 1990, et al.
(author)
-
An Uncertainty-Aware Performance Measure for Multi-Object Tracking
- 2021
-
In: IEEE Signal Processing Letters. - 1070-9908 .- 1558-2361. ; 28, s. 1689-1693
-
Journal article (peer-reviewed)abstract
- Evaluating the performance of multi-object tracking (MOT) methods is not straightforward, and existing performance measures fail to consider all the available uncertainty information in the MOT context. This can lead practitioners to select models which produce uncertainty estimates of lower quality, negatively impacting any downstream systems that rely on them. Additionally, most MOT performance measures have hyperparameters, which makes comparisons of different trackers less straightforward. We propose the use of the negative log-likelihood (NLL) of the multi-object posterior given the set of ground-truth objects as a performance measure. This measure takes into account all available uncertainty information in a sound mathematical manner without hyperparameters. We provide efficient algorithms for approximating the computation of the NLL for several common MOT algorithms, show that in some cases it decomposes and approximates the widely-used GOSPA metric, and provide several illustrative examples highlighting the advantages of the NLL in comparison to other MOT performance measures.
|
|
4. |
- Pinto, Juliano, 1990, et al.
(author)
-
Deep Learning for Model-Based Multi-Object Tracking
- 2023
-
In: IEEE Transactions on Aerospace and Electronic Systems. - : IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC. - 1557-9603 .- 0018-9251. ; 59:6, s. 7363-7379
-
Journal article (peer-reviewed)abstract
- Multi-object tracking (MOT) is the problem of tracking the state of an unknown and time-varying number of objects using noisy measurements, with important applications such as autonomous driving, tracking animal behavior, defense systems, and others. The MOT task can be divided into two settings, model-based or model-free, depending on whether accurate and tractable models of the environment are available. Model-based MOT has Bayes-optimal closed-form solutions which can achieve state-of-the-art (SOTA) performance. However, these methods require approximations in challenging scenarios to remain tractable, which impairs their performance. Deep learning (DL) methods offer a promising alternative, but existing DL models are almost exclusively designed for a model-free setting and are not easily translated to the model-based setting. This paper proposes a DL-based tracker specifically tailored to the model-based MOT setting and provides a thorough comparison to SOTA alternatives. We show that our DL-based tracker is able to match performance to the benchmarks in simple tracking tasks while outperforming the alternatives as the tasks become more challenging. These findings provide strong evidence of the applicability of DL also to the model-based setting, which we hope will foster further research in this direction.
|
|
5. |
- Pinto, Juliano, 1990
(author)
-
Deep Learning For Model-Based Multi-Object Tracking
- 2023
-
Doctoral thesis (other academic/artistic)abstract
- Multi-object tracking (MOT) is the task of estimating the state of multiple objects based on noisy sensor measurements. MOT is essential in various applications, such as pedestrian monitoring, vehicle tracking, animal behavior analysis, and others. It can be broadly divided into two categories: model-free and model-based, depending on whether accurate and tractable models of the measurement sensor and objects' dynamics are available for methods to use. In model-based MOT, closed-form, Bayes-optimal solutions can be derived for certain model families. These solutions achieve the best possible performance in expectation, but become intractable as the time-horizon increases due to an exponential growth in the number of terms. Approximations are necessary to make these methods feasible, but they result in performance degradation for challenging tracking tasks. The main objective of this thesis is to use deep learning (DL) to address this limitation. The approach taken is to treat MOT as a sequence-to-sequence learning task, devising methods that learn to map measurement sequences to state estimates directly. This perspective frees methods from the need to explicitly consider all possible associations between objects and measurements, thereby side-stepping the intractability of traditional approaches. Furthermore, the available models of the environment are leveraged to generate unlimited synthetic data. This is used to train modern DL architectures that excel in the regime of big data, unlocking their power to reason about complicated and long-term temporal interactions in their inputs. When developing the aforementioned methods, it became necessary to compare their predictions and estimated uncertainties to the state-of-the-art trackers for the model-based setting. To allow for this, another contribution of this thesis is with the paper "An Uncertainty-Aware Performance Measure for Multi-Object Tracking", which proposes the first uncertainty-aware, hyperparameter-free, mathematically principled performance measure for MOT.
|
|
6. |
- Pinto, Juliano, 1990, et al.
(author)
-
Next Generation Multitarget Trackers: Random Finite Set Methods vs Transformer-based Deep Learning
- 2021
-
In: Proceedings of 2021 IEEE 24th International Conference on Information Fusion, FUSION 2021. ; , s. 1059-1066
-
Conference paper (peer-reviewed)abstract
- Multitarget Tracking (MTT) is the problem of tracking the states of an unknown number of objects using noisy measurements, with important applications to autonomous driving, surveillance, robotics, and others. In the model-based Bayesian setting, there are conjugate priors that enable us to express the multi-object posterior in closed form, which could theoretically provide Bayes-optimal estimates. However, the posterior involves a super-exponential growth of the number of hypotheses over time, forcing state-of-the-art methods to resort to approximations for remaining tractable, which can impact their performance in complex scenarios. Model-free methods based on deep-learning provide an attractive alternative, as they can, in principle, learn the optimal filter from data, but to the best of our knowledge were never compared to current state-of-the-art Bayesian filters, specially not in contexts where accurate models are available. In this paper, we propose a high-performing deeplearning method for MTT based on the Transformer architecture and compare it to two state-of-the-art Bayesian filters, in a setting where we assume the correct model is provided. Although this gives an edge to the model-based filters, it also allows us to generate unlimited training data. We show that the proposed model outperforms state-of-the-art Bayesian filters in complex scenarios, while matching their performance in simpler cases, which validates the applicability of deep-learning also in the model-based regime. The code for all our implementations is made available at https://github.com/JulianoLagana/MT3.
|
|
7. |
- Tranheden, Wilhelm, et al.
(author)
-
DACS: Domain adaptation via cross-domain mixed sampling
- 2021
-
In: Proceedings - 2021 IEEE Winter Conference on Applications of Computer Vision, WACV 2021. ; , s. 1378-1388
-
Conference paper (peer-reviewed)abstract
- Semantic segmentation models based on convolutional neural networks have recently displayed remarkable performance for a multitude of applications. However, these models typically do not generalize well when applied on new domains, especially when going from synthetic to real data. In this paper we address the problem of unsupervised do-main adaptation (UDA), which attempts to train on labelled data from one domain (source domain), and simultaneously learn from unlabelled data in the domain of interest (target domain). Existing methods have seen success by training on pseudo-labels for these unlabelled images. Multiple techniques have been proposed to mitigate low-quality pseudo-labels arising from the domain shift, with varying degrees of success. We propose DACS: Domain Adaptation via Cross-domain mixed Sampling, which mixes images from the two domains along with the corresponding labels and pseudo-labels. These mixed samples are then trained on, in addition to the labelled data itself. We demonstrate the effectiveness of our solution by achieving state-of-the-art results for GTA5 to Cityscapes, a common synthetic-to-real semantic segmentation benchmark for UDA.
|
|