Algorithm 1 Mask-PNet Mask Logits Updating

/* For better understanding, some tensor sizes are as below*/

/* semantic: [ $H, W,C_1+C_2$ ], stuff: [ $H, W,C_1$ ], things: [ $H, W,C_2$ ]*/

/* logits: [ $H, W,C_1+k$ ], empty: [ $H, W$ ], seed_map: [h, w]*/ /* new_logits: [ $H, W,k_i$ ], feat:[h, w, c], filtered_feat:[h,w, c]*/

/* proposals: [ $k_i, c$ ], where H = W = 128 */

Input: Features $feat$ with $sizes$ in [(128,128), (64, 64), …, (2,2)];                     Semantic segmentation results: $stuff$ and $things$.         Output: Panoptic Mask Logits: $logits$.

1    $logits\leftarrow stuff$;

2      for $i= 1 to 7$ do     3                    $\left(h, w\right)\leftarrow sizes\left[i\right]$;     4                    $empty\leftarrow things>logits$;

5                    $seed\_map\leftarrow AvgLowPass\left(empty, h,w\right)>0.5$;

6                    if    $seed\_map.sum >0$ do

7                                    break;     8                    $filtered\_feat\leftarrow feat*seed\_map$;     9                    $proposals =MaskPyramid\left(filtered\_feat\right)$;   10                      $new\_logits =CNN\left(feat, proposals\right)$;   11                      $logits\leftarrow Concatenate\left(logits, new\_logits\right)$;                                             /* Return panoptic segmentation mask logits.*/   12        return $logits$;