Parallel Pipeline for Adaptive Synthetic Data Augmentationin Thermal Imaging Pose Estimation

In thermal imaging, the estimation of human poses is a challenge because widelyadopted models, like OpenPose [1], show a reduction in performance when applied to
this domain. This work proposes a data augmentation pipeline with adaptive input that
aims to increase the detection of the pre-trained human pose models on the thermal
domain. The pipeline relies on the iterative generation of synthetic thermal images,
produced by a pretrained Generative Adversarial Network (GAN), in which complete
human poses are clearly identifiable. These synthetic samples are used to improve the
training process of pose estimation models by enriching the training set with thermally
consistent, pose-preserving examples [2, 3, 4, 5].
A key contribution of this study is the design of a hybrid parallel execution pipeline,
optimized for heterogeneous high-performance computing environments [6]. This processing model works in a parallel and asynchronous manner: the GPU handles the
training of the GAN network and the synthetic image generation through the pretrained
model, while the pose detection in synthetic images is carried out using the remaining
available CPU cores.
In each iteration, the model pretrained in the previous iteration generates synthetic
images, which are then analyzed with OpenPose to identify complete human poses.
Simultaneously, the model is re-trained using those synthetic images generated in previous iterations, where at least one complete pose was detected. The number of images
generated per iteration is variable and depends on the computation time required for
training the model.
It is worth highlighting that, due to the asynchronous nature of the proposed
pipeline, the number of synthetic images generated per iteration varies according to
the system’s processing capacity. This variability directly impacts the accumulated volume of synthetic data with detected poses, the magnitude of which can shift the model’s
optimal learning point to later iterations. Experimental results indicate that, without
the use of synthetic data, training tends to require a greater number of iterations to
achieve satisfactory convergence. In contrast, the progressive incorporation of synthetic
images during training enables the model to reach this optimal point at earlier stages,
thereby improving the efficiency of the learning process.

Видео Parallel Pipeline for Adaptive Synthetic Data Augmentationin Thermal Imaging Pose Estimation канала CMMSE: Computational & Mathematics Methods