Jet Single Shot Detection
- Pol, Adrian Alan et al
- arXiv:2105.05785
 | The architecture of the SSD network, proposed by~\cite{liu2016ssd}. |
 | Energy deposits in CMS ECAL and HCAL translated to a two-dimensional image, an example input to the SSD network. The red bounding boxes correspond to ground truth with target label and momentum. |
 | Energy deposits in CMS ECAL and HCAL translated to a two-dimensional image, an example input to the SSD network. The red bounding boxes correspond to ground truth with target label and momentum. |
 | Jet-SSD architecture: input (in blue), convolution block, i.e. convolution layer followed by batch normalization and PReLU activation (in yellow), average pooling (in red) and output layer (in green). The numbers indicate the number of output channels in each block. The part of the network highlighted by dashed lines is used only during the training step. |
 | An example of the Jet-SSD at inference for one event with the input calorimeter image and highlighted true labels (top) and prediction bounding boxes (bottom). The fractional number next to the categorical label corresponds to the network confidence score. |
 | An example of the Jet-SSD at inference for one event with the input calorimeter image and highlighted true labels (top) and prediction bounding boxes (bottom). The fractional number next to the categorical label corresponds to the network confidence score. |
 | Jet-SSD PR curve and AP values for three target categories (b, W-H and t) and two weights precision (FPN and TWN). |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |
 | The mean ($\mu$) and median (Q2) localization error in $\eta$ (left column), $\phi$ (centre column) and relative error in mass regression between ground truth (GT) and Jet-SSD output (SSD) for FPN and TWN versions. The results are reported for each class independently: b jets (top row), W-H jets (centre row) and t jest (bottom row). All results are calculated as a function of p$_T$. |