.png){kind=icon}
.png){kind=icon}
Return
Encoder-decoder assisted image generation for person re-identification
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-022-11907-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Due to the low number of pedestrian samples in the categories in person Re-Identification (ReID) benchmarks, many researchers use Generative Adversarial Networks (GANs) to generate samples and expand the datasets. Real and generated samples are then used to train the person ReID model. In traditional GANs, high-dimensional samples are generated from noise. However, due to the complexity of pedestrian samples, the visual effect of generated samples is unsatisfactory. In this work, we propose a new generative model called the Encoder-Decoder Assisted Image Generative Adversarial Network (EDAGAN). EDAGAN improves the visual effects of the generated samples by reducing the dimensions of generated feature, which are obtained by the traditional GANs. In addition, many existing methods cannot optimize the real and generated samples simultaneously. Thus, the person ReID model may not make good use of the generated samples to improve the performance. For this purpose, we propose a new loss function called Soft Label Smoothing Regularization for Outliers (SLSRO), which facilitates the use of real samples and generated samples for model training. We use ResNet-50 as the backbone network to evaluate the effectiveness of EDAGAN and SLSRO. The experiments show that the EDAGAN with the SLSRO achieves a significant improvement compared to other models on the three public benchmarks, Market-1501, DukeMTMC-ReID and CUHK03.
Keywords:
Person ReID
GAN
AutoEncoder
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W