Table 2 Performance comparisons of the proposed method with other models. “Mask” and “Box” indicate using mask annotations and box annotations as ground truth, respectively. The numbers presented in the table are derived from the experimental results of the corresponding models on the test dataset. The bold fonts indicate the best results.
From: Weakly supervised learning through box annotations for pig instance segmentation
Models | Label | Box detection (%) | Mask segmentation (%) | ||||||
|---|---|---|---|---|---|---|---|---|---|
\(\textrm{AP}\) | \(\textrm{AP}_{0.5}\) | \(\textrm{AP}_{0.75}\) | \(\textrm{AP}_l\) | \(\textrm{AP}\) | \(\textrm{AP}_{0.5}\) | \(\textrm{AP}_{0.75}\) | \(\textrm{AP}_l\) | ||
Mask R-CNN | Mask | 85.41 | 96.41 | 93.02 | 86.80 | 82.11 | 96.46 | 94.28 | 83.78 |
CondInst | Mask | 86.22 | 96.30 | 92.85 | 88.73 | 86.32 | 96.48 | 95.12 | 89.74 |
YOLACT | Mask | 80.61 | 95.92 | 91.83 | 82.73 | 82.60 | 95.84 | 93.42 | 86.11 |
SOLO | Mask | - | - | - | - | 83.41 | 98.11 | 95.92 | 84.92 |
SOLOv2 | Mask | - | - | - | - | 88.51 | 96.72 | 95.60 | 90.22 |
BoxInst | Box | 85.36 | 97.84 | 93.90 | 86.67 | 81.54 | 97.96 | 94.41 | 83.91 |
DiscoBox | Box | 81.41 | 95.20 | 87.23 | 83.12 | 73.42 | 96.21 | 85.90 | 74.81 |
Ours | Box | 89.22 | 98.24 | 95.76 | 90.10 | 85.29 | 98.32 | 96.69 | 86.90 |
