# Stochastic-YOLO: Efficient Probabilistic Object Detection under Dataset Shifts

**Authors:** Tiago Azevedo, René de Jong, Partha Maji  
**arXiv:** [2009.02967](https://arxiv.org/abs/2009.02967)  
**Date:** 2020-09-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2009.02967v1/

## Key points

- Paper title: "Stochastic-YOLO: Efficient Probabilistic Object Detection under Dataset Shifts"
- Addresses challenges of uncertainty estimation and evaluation in Object Detection (OD) tasks
- Proposes Stochastic-YOLO architecture as a solution
- Stochastic-YOLO adapts YOLOv3 architecture
- Introduces stochasticity through Monte Carlo Dropout (MC-Drop)
- Generates uncertainty estimations using MC-Drop
- Efficient implementation provided to reduce burden of MC-Drop sampling during inference time
- Evaluates architecture across different levels of dataset shift
- Sensitivity analyses conducted to demonstrate improvements in uncertainty estimations, particularly spatial uncertainties
- Contributes to the field by addressing limitations in uncertainty estimation and evaluation for OD tasks
- Effective approach for generating uncertainty estimations
- Shows promise in improving spatial uncertainties

## Summary

The paper titled "Stochastic-YOLO: Efficient Probabilistic Object Detection under Dataset Shifts" addresses the challenges of uncertainty estimation and evaluation in Object Detection (OD) tasks. To overcome these challenges, the authors propose Stochastic-YOLO, a novel architecture that adapts the well-established YOLOv3 architecture. Stochastic-YOLO introduces stochasticity through Monte Carlo Dropout (MC-Drop) to generate uncertainty estimations. The architecture is evaluated across different levels of dataset shift and an efficient implementation is provided to reduce the burden of the MC-Drop sampling mechanism during inference time. Sensitivity analyses are conducted to demonstrate that Stochastic-YOLO improves different components of uncertainty estimations, particularly spatial uncertainties. This paper contributes to the field by addressing the limitations in uncertainty estimation and evaluation for OD tasks with an effective approach for generating uncertainty estimations and showing promise in improving spatial uncertainties.