In semantic segmentation, most existing real-time deep models trained with each frame independently may produce inconsistent results for a video sequence. Advanced methods take into considerations the correlations in the video sequence, e.g., by propagating the results to the neighboring frames using optical flow, or extracting the frame representations with other frames, which may lead to inaccurate results or unbalanced latency. In this work, we process efficient semantic video segmentation in a per-frame fashion during the inference process.

Different from previous per-frame models, we explicitly consider the temporal consistency among frames as extra constraints during the training process and embed the temporal consistency into the segmentation network. Therefore, in the inference process, we can process each frame independently with no latency, and improve the temporal consistency with no extra computational cost and post-processing. We employ compact models for real-time execution. To narrow the performance gap between compact models and large models, new knowledge distillation methods are designed. Our results outperform previous keyframe based methods with a better trade-off between the accuracy and the inference speed on popular benchmarks, including the Cityscapes and Camvid. The temporal consistency is also improved compared with corresponding baselines which are trained with each frame independently.

This repository contains the demo evaluate code and the training scripts for motion loss of our paper (ECCV2020) Efficient Semantic Video Segmentation with Per-frame Inference.

We generated new pseudo labels with test-time augmentation on video sequence in data-coarse. Besides, we filter out the regions with low confidence. Training with the provided dataset, most models can achieve a better results with mIoU and temporal consistency. The pseudo data can be downloaded from this link.

Demo video for the PSPnet-18 on Cityscapes

ETC with mIoU 73.1, temporal consistncy 70.56 vs baseline with mIoU 69.79, temporal consistncy 68.50:

We employ the temporal loss the temporal knowledge distillation methods to adapte single frame image segmentation methods for semantic video segmentation methods.

Note: Other chcekpoints can be obtained by email: yifan.liu04@adelaide.edu.au if needed.

python3.5

pytorch >1.0.0

We recommend to use Anaconda.

We have tested our code on Ubuntu 16.04.

• The flownet need to be compiled following FlowNetV2
• You can first clone the FlowNetV2, and compile it. -Note that your cuda version [nvcc -V] should be the same as your torch cuda versio [torch.version.cuda]
• Then copy the folder of flownet2-pytorch/networks/resample2d_package,correlation_package,channelnorm_package to OURS/flownet/
• Download the weight of the flownet, and place it in OURS/pretrained_model/ To train the model, you also need to install apex.

1. python tool/train_with_flow.py.

1. download the Cityscape dataset
2. python tool/demo.py.

To evaluate the temporal consistency, you need to install the flownet first.

1. You need to download the video data of the cityscapes: leftImg8bit_sequence_trainvaltest.zip
2. The download data should be placed in data/cityscapes/leftImg8bit/
3. Generate the results for the sampled frames, which need to be evaluated: python tool/gen_video.py
4. Evaluate the temporal consistency based on the warpping mIoU: python tool/eval_tc.py

Note that the first time you evaluate the TC, the code will save the flow automatically.

• In our paper, we random sample ~20% of the validation set for testing the TC for all models for efficiency (lists are in 'data/list/cityscapes/val_sam').
• If you want to evaluate with all the validation video clips, you can relpace the 'data/list/cityscapes/val_video_img_sam.lst' with 'data/list/cityscapes/val_video_img.lst', and replace the 'data/list/cityscapes/val_sam' with 'data/list/cityscapes/val'. The trendency of the TC are similar.

Please change the ckpt_path in config to compare the results with baseline models

We only release the training code for the motion loss. If you are interested in the temporal knowledge distillation, you can refer to the ./utils/crit.py. All the distillation loss are included. The distillation method is a plug-in method, which can be applied to different teacher-student network pairs.

The test code borrows from semseg.

If you find this code useful, please cite:

@article{liu2020efficient,
  title={Efficient Semantic Video Segmentation with Per-frame Inference},
  author={Liu, Yifan and Shen, Chunhua and Yu, Changqian and Wang, Jingdong},
  journal={ECCV},
  year={2020}
}