Table 3.
Performance comparisons on the test set of MJU-Waste. For each method, we report the IoU for waste objects (IoU), mean IoU (mIoU), pixel Precision for waste objects (Prec) and Mean pixel precision (Mean). See Section 4.3 for details.
| Dataset: | |||||
|---|---|---|---|---|---|
| MJU-Waste (Test) | Backbone | IoU | mIoU | Prec | Mean |
| Baseline Approaches | |||||
| FCN-8s [17] | VGG-16 | 75.28 | 87.35 | 85.95 | 92.83 |
| PSPNet [21] | ResNet-101 | 78.62 | 89.06 | 86.42 | 93.11 |
| CCNet [22] | ResNet-101 | 83.44 | 91.54 | 92.92 | 96.35 |
| DeepLabv3 [23] | ResNet-50 | 79.92 | 89.73 | 86.30 | 93.06 |
| DeepLabv3 [23] | ResNet-101 | 84.11 | 91.88 | 89.69 | 94.77 |
| Proposed Multi-Level (ML) Model | |||||
| FCN-8s-ML | VGG-16 | 82.29 | 90.95 | 91.75 | 95.76 |
| (+7.01) | (+3.60) | (+5.80) | (+2.93) | ||
| PSPNet-ML | ResNet-101 | 81.81 | 90.70 | 89.65 | 94.73 |
| (+3.19) | (+1.64) | (+3.23) | (+1.62) | ||
| CCNet-ML | ResNet-101 | 86.63 | 93.17 | 96.05 | 97.92 |
| (+3.19) | (+1.63) | (+3.13) | (+1.57) | ||
| DeepLabv3-ML | ResNet-50 | 84.35 | 92.00 | 91.73 | 95.78 |
| (+4.43) | (+2.27) | (+5.43) | (+2.72) | ||
| DeepLabv3-ML | ResNet-101 | 87.84 | 93.79 | 94.43 | 97.14 |
| (+3.73) | (+1.91) | (+4.74) | (+2.37) | ||