PaperlayerView exact revision on arXiv

Source-linked AI summary

Monte Carlo DropBlock for Modelling Uncertainty in Object Detection

Kumari Deepshikha, Sai Harsha Yelleni, P. K. Srijith, C Krishna Mohan

arXiv:2108.03614v1cs.CV

TL;DR

High-risk vision applications need uncertainty estimates for out-of-distribution predictions, but standard models can be overconfident. The paper applies MC-DropBlock during training and testing, and reports improved generalization, calibration, and uncertainty modeling in YOLO detection and segmentation.

  • Problem

    Standard deep-learning vision models can make erroneous, overconfident predictions on out-of-distribution data, a concern in high-risk applications.

    Source: §1 · pp. 1–2

  • Method

    MC-DropBlock applies DropBlock during training and testing in convolutional models, with Gaussian likelihoods also used to model aleatoric uncertainty.

    Source: §1 · p. 3

  • Results

    MC-DropBlock improves YOLO generalization, calibration, and uncertainty modeling for object detection and segmentation.

    Source: §1 · p. 3, §6 · p. 14

  • Takeaways & Limitations

    The authors conclude MC-DropBlock is better suited than MC-Dropout for uncertainty modeling in fully convolutional architectures.

    Source: §6 · p. 14

Abstract

from arXiv · show

With the advancements made in deep learning, computer vision problems like object detection and segmentation have seen a great improvement in performance. However, in many real-world applications such as autonomous driving vehicles, the risk associated with incorrect predictions of objects is very high. Standard deep learning models for object detection such as YOLO models are often overconfident in their predictions and do not take into account the uncertainty in predictions on out-of-distribution data. In this work, we propose an efficient and effective approach to model uncertainty in object detection and segmentation tasks using Monte-Carlo DropBlock (MC-DropBlock) based inference. The proposed approach applies drop-block during training time and test time on the convolutional layer of the deep learning models such as YOLO. We show that this leads to a Bayesian convolutional neural network capable of capturing the epistemic uncertainty in the model. Additionally, we capture the aleatoric uncertainty using a Gaussian likelihood. We demonstrate the effectiveness of the proposed approach on modeling uncertainty in object detection and segmentation tasks using out-of-distribution experiments. Experimental results show that MC-DropBlock improves the generalization, calibration, and uncertainty modeling capabilities of YOLO models in object detection and segmentation.

NVIDIA India

The passage identifies P.K. Srijith with IIT Hyderabad.

1 Introduction

The introduction argues that high-risk vision applications need reliable uncertainty estimates, especially on out-of-distribution data. It proposes MC-DropBlock for convolutional models and reports improvements across uncertainty modeling, calibration, and generalization.

1 Introduction

  • High-risk applications face out-of-distribution inputs, where models may make erroneous predictions with high confidence.The introduction names autonomous driving and healthcare as examples.Source: §1 · pp. 1–2
  • DropBlock drops contiguous neuron regions, addressing dropout's failure to account for spatial collocation in convolutional networks.Source: §1 · p. 3
  • MC-DropBlock applies DropBlock during training and testing, using multiple architectures to model uncertainty in convolutional vision models.The approach is presented for YOLO object detection and image segmentation.Source: §1 · p. 3
  • The paper theoretically equates MC-DropBlock with variational inference on a Bayesian convolutional neural network.Source: §1 · p. 3
  • The authors report that MC-DropBlock improves YOLO generalization, calibration, and uncertainty modeling for object detection and segmentation.Source: §1 · p. 3

2 Related Work

Prior work addresses uncertainty in vision models using Bayesian methods, dropout, and Gaussian prediction heads; the authors motivate DropBlock for convolutional networks.

2 Related Work

  • Bayesian neural networks model uncertainty through parameter posteriors, while Monte Carlo dropout offers a practical approximation by applying dropout at training and test time.Source: §2 · p. 3
  • DropBlock was introduced to address dropout's ineffectiveness on convolutional networks by dropping contiguous neuron regions.Source: §2 · p. 4
  • Prior detection work modeled aleatoric uncertainty with Gaussian YOLOv3 and epistemic uncertainty with MC dropout, which the authors describe as ineffective for convolutional networks.The paper proposes Monte Carlo DropBlock as an alternative for uncertainty modeling and generalization in convolutional networks.Source: §2 · p. 4

3 Background

The background introduces Bayesian prediction and Monte Carlo dropout as context for the paper's convolutional detection and segmentation models.

3 Background

  • The background frames supervised learning as mapping input examples to outputs through linear-nonlinear transformations and convolution operations.Source: §3 · p. 4

3.1 Bayesian Neural Networks

  • Bayesian neural networks place distributions over model parameters and estimate a posterior from observed data.Predictions average outputs from multiple posterior weight samples, a procedure called Bayesian model averaging.Source: §3.1 · pp. 4–5

3.2 Monte Carlo Dropout

  • Monte Carlo dropout applies dropout at test time and averages predictions across sampled dropout architectures to estimate epistemic uncertainty.Source: §3.2 · pp. 5–6

3.3 Object Detection and Segmentation

  • YOLO uses convolutional layers for object detection, while YOLACT combines prototype-mask generation with per-instance mask-coefficient prediction for real-time segmentation.Source: §3.3 · p. 6

4 Monte Carlo Drop Block

The section presents MC-DropBlock as structured masking at training and inference to model epistemic uncertainty, then combines it with Gaussian box modeling to capture aleatoric uncertainty.

4 Monte Carlo Drop Block

  • MC-DropBlock applies contiguous block masks during training and inference, averaging predictions across an exponential ensemble of subnetworks to model epistemic uncertainty.DropBlock masks contiguous regions of feature maps, unlike independently sampled masks.Source: §4 · p. 7, Figure 2 · p. 7

4.1 Drop block as Bayesian Approximation

  • Each filter block is either zero with drop probability γ or W with probability 1−γ, defining a structured variational posterior for the Bayesian approximation.Source: §4.1 · p. 9

4.2 Modeling Aleatoric uncertainty

  • Gaussian YOLO models four box coordinates with means and variances (8 outputs rather than 4); combined with MC-DropBlock and GIoU, it models aleatoric and epistemic uncertainty.Source: §4.2 · p. 9

5 Experiments

The experiments assess MC-DropBlock across out-of-distribution uncertainty, object detection, segmentation, and classification, comparing it with baseline models and alternative DropBlock or dropout approaches. Results report stronger OOD entropy, competitive detection and segmentation performance, and improved classification accuracy, alongside qualitative examples where MC-DropBlock avoids false detections.

5 Experiments

  • DropBlock was applied in later model layers; experiments compared training-time, inference-time, and MC-DropBlock across OOD uncertainty, calibration, and task performance.The authors report that applying DropBlock in initial layers significantly reduced performance; they tested several block sizes and drop probabilities.Source: §5 · pp. 10–11, §5.2 · p. 12

5.1 Out-of-Distribution Experiments

  • MC-DropBlock produced the highest OOD entropy across YOLO models and ResNets, outperforming MC-Dropout and other DropBlock approaches in object detection and classification.The object-detection OOD test trained on Pascal VOC and evaluated on COCO.Source: §5.1 · pp. 11–12, Table 1 · p. 11

5.2 Object Detection task

  • MC-DropBlock ranked second in COCO detection mAP behind training-time DropBlock and second in Brier calibration, ahead of baseline and training-only DropBlock.Adding Gaussian loss to model aleatoric uncertainty improved YOLO mAP.Source: §5.2 · p. 12, Table 2 · p. 12, Table 3 · p. 12

5.3 Segmentation task

  • 1.1% on mask mAP and 1.05% on box mAP: inference-time DropBlock improved YOLACT-550, while MC-DropBlock also beat the baseline.The experiments used the COCO segmentation dataset.Source: §5.3 · p. 13, Table 4 · p. 13

5.4 Classification task

  • 92.98% top-1 accuracy using MC-DropBlock, versus 92.26% using MC Dropout and 92.08% using baseline ResNet-110 on CIFAR-10.On the OOD image-classification test, MC-DropBlock also had the highest entropy.Source: §5.4 · p. 13

5.5 Qualitative Analysis

  • MC-DropBlock removed false detections in two examples: pot misclassified as aeroplanes and bread misclassified as sheep by baseline YOLOv5 and MC-Dropout.The examples used models trained on Pascal VOC and images from Open Images Data.Source: §5.5 · p. 13, Figure 5 · p. 14, Figure 6 · p. 14

6 Conclusion

MC-DropBlock models uncertainty in convolutional vision networks, and the authors report improved calibration, uncertainty modeling, and generalization when DropBlock is applied at test time.

  • The authors report that test-time DropBlock improves calibration, uncertainty modeling, and generalization performance.Source: §6 · p. 14
  • MC-DropBlock can be viewed as variational inference on a Bayesian convolutional neural network.Source: §6 · p. 14
  • Gaussian loss is incorporated in YOLO models to handle aleatoric uncertainty.Source: §6 · p. 14
  • For fully convolutional layer-based architectures, the authors describe MC-DropBlock as better than MC Dropout for modeling uncertainty.Source: §6 · p. 14
  • The authors report testing the method on multiple vision tasks and suggest fully convolutional layer-based tasks could benefit from it.Source: §6 · p. 14
Loading 2108.03614v1…