https://github.com/linksense/LightNetPlusPlus
LightNet++: Boosted Light-weighted Networks for Real-time Semantic Segmentation
https://github.com/linksense/LightNetPlusPlus
apex aspp attention-mechanism bifpn cityscapes deepdrive deeplab-v3-plus dense-aspp efficientnet inplaceabn light-weighted-network lightnet mixnet mobilenetv2 mobilenetv2plus pytorch semantic-segmentation shufflenetv2 shufflenetv2plus unsharp-masking
Last synced: 9 months ago
JSON representation
LightNet++: Boosted Light-weighted Networks for Real-time Semantic Segmentation
- Host: GitHub
- URL: https://github.com/linksense/LightNetPlusPlus
- Owner: linksense
- License: mit
- Created: 2019-02-27T02:59:04.000Z (about 7 years ago)
- Default Branch: master
- Last Pushed: 2020-03-24T10:42:58.000Z (about 6 years ago)
- Last Synced: 2024-05-15T04:39:38.687Z (almost 2 years ago)
- Topics: apex, aspp, attention-mechanism, bifpn, cityscapes, deepdrive, deeplab-v3-plus, dense-aspp, efficientnet, inplaceabn, light-weighted-network, lightnet, mixnet, mobilenetv2, mobilenetv2plus, pytorch, semantic-segmentation, shufflenetv2, shufflenetv2plus, unsharp-masking
- Language: Python
- Homepage: https://ansleliu.github.io/LightNet.html
- Size: 64.6 MB
- Stars: 238
- Watchers: 15
- Forks: 54
- Open Issues: 5
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# LightNet++
## !!!New Repo.!!! ⇒ **[EfficientNet.PyTorch: Concise, Modular, Human-friendly PyTorch implementation of EfficientNet with Pre-trained Weights](https://github.com/ansleliu/EfficientNet.PyTorch)**
## !!!New Repo.!!! ⇒ **[MixNet-Pytorch: Concise, Modular, Human-friendly PyTorch implementation of MixNet with Pre-trained Weights](https://github.com/ansleliu/MixNet-Pytorch)**
This repository contains the code (PyTorch-1.0+, **W.I.P.**) for: "**LightNet++: Boosted Light-weighted Networks for Real-time Semantic Segmentation**" by Huijun Liu.
**LightNet++** is an advanced version of **[LightNet](https://github.com/ansleliu/LightNet)**, which purpose to get more concise model design,
smaller models, and better performance.
- **MobileNetV2Plus**: Modified MobileNetV2 (backbone)[[1,8]](#references) + DSASPPInPlaceABNBlock[[2,3]](#references) +
Parallel Bottleneck Channel-Spatial Attention Block (PBCSABlock)[[6]](#references) + UnSharp Masking (USM) + Encoder-Decoder Arch.[[3]](#references) +
InplaceABN[[4]](#references).
- **ShuffleNetV2Plus**: Modified ShuffleNetV2 (backbone)[[1,8]](#references) + DSASPPInPlaceABNBlock[[2,3]](#references) +
Parallel Bottleneck Channel-Spatial Attention Block (PBCSABlock)[[6]](#references)+ UnSharp Masking (USM) + Encoder-Decoder Arch.[[3]](#references) +
InplaceABN[[4]](#references).
- [**MixSeg-MixBiFPN**](https://github.com/ansleliu/LightNetPlusPlus/blob/master/models/mixnetseg.py): Modified MixNet (backbone)[[1,8]](#references) + MixBiFPNBlock[[2,3]](#references) + Encoder-Decoder Arch.[[3]](#references)
More about **USM(Unsharp Mask)-Operator Block** see Repo: [**SharpPeleeNet**](https://github.com/ansleliu/SharpPeleeNet)
## Dependencies
- [Python3.6](https://www.python.org/downloads/)
- [PyTorch(1.0.1+)](http://pytorch.org)
- [inplace_abn](https://github.com/mapillary/inplace_abn)
- [apex](https://github.com/NVIDIA/apex): Tools for easy mixed precision and distributed training in Pytorch
- [tensorboard](https://www.tensorflow.org/programmers_guide/summaries_and_tensorboard)
- [tensorboardX](https://github.com/lanpa/tensorboard-pytorch)
- [tqdm](https://github.com/tqdm/tqdm)
### Datasets for Autonomous Driving
- [Cityscapes](https://www.cityscapes-dataset.com/)
- [Mapillary Vistas Dataset](https://www.mapillary.com/dataset/vistas)
- [Berkeley Deep Drive (BDD100K)](https://bdd-data.berkeley.edu/)
- [ApolloScape](http://apolloscape.auto/index.html#)
## Results
### Results on Cityscapes (Pixel-level/Semantic Segmentation)
| Model | mIoU (S.S* Mixed Precision) |Model Weight|
|---|---|---|
|**MobileNetV2Plus X1.0**|[71.5314 (**WIP**)](https://github.com/ansleliu/LightNetPlusPlus/blob/master/checkpoint/MobileNetv2Plus.csv)|[cityscapes_mobilenetv2plus_x1.0.pkl (14.3 MB)](https://github.com/ansleliu/LightNetPlusPlus/blob/master/checkpoint/cityscapes_mobilenetv2plus_x1.0.pkl)|
|**ShuffleNetV2Plus X1.0**|[69.0885-72.5255 (**WIP**)](https://github.com/ansleliu/LightNetPlusPlus/blob/master/checkpoint/ShuffleNetV2PlusX1.0.csv)|[cityscapes_shufflenetv2plus_x1.0.pkl (8.59 MB)](https://github.com/ansleliu/LightNetPlusPlus/blob/master/checkpoint/cityscapes_shufflenetv2plus_x1.0.pkl)|
|**MixSeg+MixBiFPN ArchS**|[72.2321 (**WIP**)](https://github.com/ansleliu/LightNetPlusPlus/blob/master/checkpoint/MixSegMixBiFPN_ArchS.csv)|[cityscapes_mixseg_archs_mixbifpn.pkl (16.4 MB)](https://github.com/ansleliu/LightNetPlusPlus/blob/master/checkpoint/cityscapes_mixseg_archs_mixbifpn.pkl)|
* S.S.: Single Scale (1024x2048)
## Feature Visualization
