Abstract:
"In snowy conditions, there is a lack of visibility for Bird’s Eye View Detectors, impacting the accuracy of the BEV Detectors. This issue is critical in applications like autonomous vehicles, where accurate environment visibility is crucial for a safer navigation. The existing solutions often fall short in these challenging weather scenarios, requiring the exploration of other alternative methods to enhance the input image for the BEV Detectors. This thesis aims to address this problem using a GAN-based snow removal model.
In this Thesis, GANs are used to enhance the raw input data by effectively reducing noise and enhancing visibility in snowy weather conditions. The proposed methodology involves a 2-step process; first, the snowy input image is captured and passed through our GAN model, and then the improved, snow removed image are fed into the existing BEV Detectors (eg. FCOS3D) in MMDETECTION3D. This approach ensures that the BEV Detectors receive clear input images, ultimately leading to improved detection performance in snowy weather conditions.
Using the Pix2Pix GAN architecture as its foundation, the GAN demonstrated proficiency in snow removal from images. The conducted experiments indicate that the model excels in this task, achieving a PSNR score of up to 31.1941 dB and SSIM scores reaching as high as 0.9216, following 65 epochs of training on the dataset."
