Blackout - an augmentation technique for occlusion handling in pedestrian detection
Pedestrian Detection has numerous applications, ranging from autonomous driving to traffic analysis. One inherent challenge faced in the field of pedestrian detection is the problem of occluded objects. In this project, we propose Blackout - a novel enhancement to an existing data augmentation Cutou...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2023
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/166758 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
id |
sg-ntu-dr.10356-166758 |
---|---|
record_format |
dspace |
spelling |
sg-ntu-dr.10356-1667582023-05-12T15:36:45Z Blackout - an augmentation technique for occlusion handling in pedestrian detection Satini, Sankeerthana Vidya Sudarshan School of Computer Science and Engineering vidya.sudarshan@ntu.edu.sg Engineering::Computer science and engineering Pedestrian Detection has numerous applications, ranging from autonomous driving to traffic analysis. One inherent challenge faced in the field of pedestrian detection is the problem of occluded objects. In this project, we propose Blackout - a novel enhancement to an existing data augmentation Cutout, which tackles intraclass occlusion, where a pedestrian is blocked by another pedestrian. Blackout is a strong augmentation technique that blacks out the intersection of overlapping ground truth bounding boxes. It does so by computing the intersection over union of the ground truth bounding boxes present in an image and blacks out the intersection of the ground truth bounding boxes if the intersection over union values lies within the threshold range. Blackout has been evaluated on the Penn Fudan Pedestrian Detection Dataset. Through our evaluation of Blackout, it can be concluded that Blackout implemented as a standalone strong augmentation technique on larger models can result in a large increase in accuracy – a 13% increase in accuracy in the case of the state-of-the-art model You Only Live Once eXtreme – Medium sized model (YOLOX-M). Blackout can also be used as an additional augmentation along with other existing augmentations as it results in a 1% and 2% increase in accuracy in YOLOX-S and YOLOX-M respectively. Therefore, Blackout can prove to be extremely useful in cases where we would want to force the model to learn from the outlines of the objects. This can include a myriad of applications in field such as crowd tracking and control. Bachelor of Engineering (Computer Science) 2023-05-12T07:12:11Z 2023-05-12T07:12:11Z 2023 Final Year Project (FYP) Satini, S. (2023). Blackout - an augmentation technique for occlusion handling in pedestrian detection. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/166758 https://hdl.handle.net/10356/166758 en SCSE22-0486 application/pdf Nanyang Technological University |
institution |
Nanyang Technological University |
building |
NTU Library |
continent |
Asia |
country |
Singapore Singapore |
content_provider |
NTU Library |
collection |
DR-NTU |
language |
English |
topic |
Engineering::Computer science and engineering |
spellingShingle |
Engineering::Computer science and engineering Satini, Sankeerthana Blackout - an augmentation technique for occlusion handling in pedestrian detection |
description |
Pedestrian Detection has numerous applications, ranging from autonomous driving to traffic analysis. One inherent challenge faced in the field of pedestrian detection is the problem of occluded objects. In this project, we propose Blackout - a novel enhancement to an existing data augmentation Cutout, which tackles intraclass occlusion, where a pedestrian is blocked by another pedestrian.
Blackout is a strong augmentation technique that blacks out the intersection of overlapping ground truth bounding boxes. It does so by computing the intersection over union of the ground truth bounding boxes present in an image and blacks out the intersection of the ground truth bounding boxes if the intersection over union values lies within the threshold range. Blackout has been evaluated on the Penn Fudan Pedestrian Detection Dataset.
Through our evaluation of Blackout, it can be concluded that Blackout implemented as a standalone strong augmentation technique on larger models can result in a large increase in accuracy – a 13% increase in accuracy in the case of the state-of-the-art model You Only Live Once eXtreme – Medium sized model (YOLOX-M). Blackout can also be used as an additional augmentation along with other existing augmentations as it results in a 1% and 2% increase in accuracy in YOLOX-S and YOLOX-M respectively.
Therefore, Blackout can prove to be extremely useful in cases where we would want to force the model to learn from the outlines of the objects. This can include a myriad of applications in field such as crowd tracking and control. |
author2 |
Vidya Sudarshan |
author_facet |
Vidya Sudarshan Satini, Sankeerthana |
format |
Final Year Project |
author |
Satini, Sankeerthana |
author_sort |
Satini, Sankeerthana |
title |
Blackout - an augmentation technique for occlusion handling in pedestrian detection |
title_short |
Blackout - an augmentation technique for occlusion handling in pedestrian detection |
title_full |
Blackout - an augmentation technique for occlusion handling in pedestrian detection |
title_fullStr |
Blackout - an augmentation technique for occlusion handling in pedestrian detection |
title_full_unstemmed |
Blackout - an augmentation technique for occlusion handling in pedestrian detection |
title_sort |
blackout - an augmentation technique for occlusion handling in pedestrian detection |
publisher |
Nanyang Technological University |
publishDate |
2023 |
url |
https://hdl.handle.net/10356/166758 |
_version_ |
1770565401719078912 |