Fig. 1. High frequency and low frequency of feature map
Fig. 2. Schematic of OctConv operation
Fig. 3. Comparison between swish and ReLU
Fig. 4. Comparison between swish function and hard-swish function
Fig. 5. OTCH-L module
Fig. 6. Proposed model
Fig. 7. Schematic of MIoU
Fig. 8. Influence of α on network performance
Fig. 9. Schematic of residual modules with different dimensionality reduction channels
Fig. 10. Comparison of experimental results
Category | Content |
---|
Person | person | Animal | bird, cat, cow, dog, horse, sheep | Vehicle | plane, bicycle, boat, bus, car, motorbike, train | Indoor | bottle, chair, dining table, potted plant, sofa, tv /monitor |
|
Table 1. Categories contained in the dataset
Parameter | Value |
---|
Batch size | 8 | Input image size | 473×473 | Epoch | 100 | Learning rate | 1× |
|
Table 2. Model super parameters
Network | | Accuracy /% | Speed /(frame·s-1) |
---|
ResNet-50 | | 74.5 | 8.1 | R+OctConv | 0 | 74.7 | 8.9 | R+OctConv | 0.3 | 76.2 | 11.3 | R+OctConv | 0.6 | 75.5 | 14.1 | R+OctConv | 0.9 | 73.1 | 17.5 | MobileNet v3 | | 66.1 | 15.3 | M+OctConv | 0 | 66.8 | 15.6 | M+OctConv | 0.3 | 68.3 | 19.2 | M+OctConv | 0.6 | 67.1 | 22.3 | M+OctConv | 0.9 | 65.7 | 25.4 |
|
Table 3. Influence of different α values on model performance
Module | MIoU /% | Parameter /103 | Speed /(frame·s-1) |
---|
OTCH-L(C) | 71.54 | 756 | 22.86 | OTCH-L(C/2) | 70.34 | 571 | 25.94 | OTCH-L(C/4) | 67.01 | 477 | 26.42 |
|
Table 4. Performance comparison between neck blocks with different dimensionality reduction channels
Network | MIoU /% | MPA /% | Speed /(frame·s-1) |
---|
SegNet | 59.86 | 72.45 | 10.57 | PSPNet | 65.28 | 81.51 | 16.68 | DeepLab v3 plus | 66.33 | 82.04 | 19.56 | This Paper | 70.34 | 83.65 | 25.94 |
|
Table 5. Performance comparison between different networks
Network | MIoU /% | Speed /(frame·s-1) |
---|
Other paper | 68.43 | 14.14 | This Paper | 70.34 | 25.94 |
|
Table 6. Performance comparison between the proposed network and a recent algorithm
Module | MIoU /% | MPA /% | Speed /(frame·s-1) |
---|
ResNet50 | 72.07 | 85.06 | 11.68 | Xception | 69.33 | 83.74 | 17.56 | MobileNet v3 | 67.92 | 81.39 | 26.60 | Improved MobileNet v3 | 70.34 | 83.65 | 25.94 |
|
Table 7. Performance comparison of different feature extraction networks