Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
Asthestarsfalll 073b3083ce | 1 year ago | |
---|---|---|
.. | ||
README.md | 1 year ago | |
lpsnet_l_cityscapes_1536x1024_200k.yml | 1 year ago | |
lpsnet_m_cityscapes_1536x1024_200k.yml | 1 year ago | |
lpsnet_s_cityscapes_1536x1024_200k.yml | 1 year ago |
Zhang, Yiheng and Yao, Ting and Qiu, Zhaofan and Mei, Tao. "Lightweight and Progressively-Scalable Networks for Semantic Segmentation."
arXiv preprint arXiv:2207.13600 (2022).
Model | Backbone | Resolution | Training Iters | mIoU | mIoU (flip) | mIoU (ms+flip) | Links |
---|---|---|---|---|---|---|---|
lpsnet_s | - | 1536x769 | 200000 | 71.73% | 72.71% | 73.76% | model | log | vdl |
lpsnet_m | - | 1536x769 | 200000 | 75.29% | 76.03% | 77.03% | model | log | vdl |
lpsnet_l | - | 1536x769 | 200000 | 75.72% | 76.53% | 77.23% | model | log | vdl |
Note that: Since the original paper does not provide all the implementation details, nor release the official training code, we first pre-trained the models on the ImageNet dataset, and then fine-tuned the models on the Cityscapes dataset. Specifically, compared with the experimental settings in the original paper, we made two significant changes during fine-tuning to obtain the above results:
For the pre-training results of the models on the ImageNet dataset, please see the ImageNet section, where we also have a discussion on the impact of the pre-trained model.
Model | Epoch | Top-1 accuracy | Top-5 accuracy |
---|---|---|---|
lpsnet_s | 120 | 0.403 | 0.666 |
lpsnet_m | 100 | 0.543 | 0.786 |
lpsnet_l | 120 | 0.525 | 0.773 |
Training settings are as following.
Optimizer
Optimizer | LR Scheduler | Regularizer |
---|---|---|
type: Momentum | type: Piecewice | type: L2 |
momentum: 0.9 | lr: 0.1 | coeff:0.0001 |
use_nesterov: true | decay_epochs: [30, 60, 90] | - |
- | values: [0.1, 0.01, 0.001, 0.0001] | - |
Data Augmentation
RandCropImage | RandFlipImage | Normalize |
---|---|---|
size: 224 | flip horizontal | ImageNet mean and std |
With the pre-trained model, lpsnet_m is able to obtain 74.28% mIoU on Cityscapes val set under the same experimental settings as the original paper.
Further more, we also tried more complicated data augmentation strategies, e.g. random erasing and batched mix-up. The performance of classification improved by a considerable margin: on the ImageNet dataset, the top-1 accuracy increased to 0.564 and the top-5 accuracy increased to 0.805. However, we did not observe better segmentation performance of the model on the Cityscapes dataset using these pre-trained models. Therefore, these data augmentation strategies were not used.
The paddlepaddle version of LPSNet
Python Shell Markdown Java C++ other
Dear OpenI User
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
For more agreement content, please refer to the《Openl Qizhi Community AI Collaboration Platform Usage Agreement》