.png){kind=icon}
.png){kind=icon}
Return
An instance-level data balancing method for object detection via contextual information alignment
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2024.105155.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The imbalance issues in object detection training data, such as in categories, scales, and spatial distribution, result in detection models failing to effectively fit unbalanced data. Our method aims to mitigate the performance disparity caused by data imbalance from the perspective of instance-level augmentation. Firstly, we designed a dynamic data balancing mechanism (DDBM) to develop category expansion rate, scale ratio rules, and spatial distribution indicators to alleviate data imbalance. Then, based on the pixel-level fine-grained context, a finegrained local object instance augmentation (FLOIA) method is designed to selectively copy the object instance according to the background Mosaic degree. In addition, based on the coarse-grained global context and dynamic balancing mechanism, we proposes a coarse-grained global object instance augmentation (CGOIA) method to establish an object-background association, ensure the alignment of the context information of the object instance and alleviate the data imbalance. We train the proposed instance augmentation-treated datasets on various models, resulting in improved balance across different categories and scales. Additionally, visual analysis validates that this approach balances spatial distributions while conforming to contextual information. Furthermore, this method proves advantageous for training with small-sample datasets.
Keywords:
Object detection
Data imbalance
Instance augmentation
Context alignment
Dynamic data balancing mechanism
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K