Fig. 9.

Graphical overview of the participants’ methodologies for Task 1 as described in Section 4 (Key: $X$ - input frame; $y$ - groundtruth; $\hat{y}$ - prediction). AQ-ENIB (a) proposed an ensemble of DenseNet models with Test Time Augment (TTA). BioPolimi (b) combined ResNet50 features with a Histogram of Oriented Gradients (HoG) computed on $X$. RREB (c) proposed a multi-task $U^{2}Net$ for segmentation and multi-scale regression of HoG features ($\widehat{{\mathrm{HoG}}_0}$, $\widehat{{\mathrm{HoG}}_1}$, …) computed on $y$ (${\mathrm{HoG}}_0$, ${\mathrm{HoG}}_1$, …). GRECHID (d) used 3 SEResNeXt-UNet models individually trained on each class ensembled by thresholding, where ${\mathrm{pixels}}_{\mathrm{HighConfidence}}$ are pixels predicted with high confidence and ${\mathrm{count}}_{\mathrm{threshold}}$ is the empirical threshold. SANO (e) proposed a mean ensemble of Feature Pyramid Network (FPN) with ResNet152 backbone. OOF (f) used an EfficientNet UNet++, preprocessing images with contrast-limited adaptive histogram equalization (CLAHE) and median filter.