This is the official released code for our paper, The Emergence of Objectness: Learning Zero-Shot Segmentation from Videos, and our presentation slides published at NeurIPS 2021.
@article{liu2021emergence,
title={The Emergence of Objectness: Learning Zero-Shot Segmentation from Videos},
author={Liu, Runtao and Wu, Zhirong and Yu, Stella and Lin, Stephen},
journal={Advances in Neural Information Processing Systems},
volume={34},
year={2021}
}
Our code is implemented based on the framework of MMSegmentation with PyTorch 1.5 and CUDA 10.1. Please see more environment information in the file requirements.txt.
We released our YouTube-VOS trained model which can be used for zero-shot segmentation on single images or single videos, with channel 0 for the foreground prediction.
Please download the released model and put it at the directory of this repository. data_example/ provides the example data for inference. Here is the inference command example using the released model:
PT_OUTPUT_DIR=output_test sh sh_train_pt_param.sh configs/config_test.py $GPU $port
The output results can be found in output_test/eval_test_0/.
And here is the training command example:
PT_OUTPUT_DIR=$DIR_NAME sh sh_train_pt_param.sh configs/config_train.py $GPU $port
We provide dataset format examples in the section Testing with the released model with the directory data_example/.
To train a model on a new dataset, please set the correct path in the line data_root and the line of filename split in configs/config_train.py. A split file contains a list of file names. An example of YoutubeVOS 2019 can be:
train_all_frames/JPEGImages/fa88d48a92 00000.jpg 00001.jpg 00002.jpg 00003.jpg 00004.jpg 00005.jpg
train_all_frames/JPEGImages/df59cfd91d 00000.jpg 00001.jpg 00002.jpg 00003.jpg 00004.jpg 00005.jpg
Each line represents a video frame directory and the frame files are followed. We also provide a full version train split of YoutubeVOS2019 ytb_2019.txt.
The setting of the dataset and training parameters like iterations should be set in configs/config_train.py. You could train with the following command:
PT_OUTPUT_DIR=$DIR_NAME sh sh_train_pt_param.sh configs/config.py $GPU $port
Humans can easily segment moving objects without knowing what they are. That objectness could emerge from continuous visual observations motivates us to model grouping and movement concurrently from unlabeled videos. Our premise is that a video has different views of the same scene related by moving components, and the right region segmentation and region flow would allow mutual view synthesis which can be checked from the data itself without any external supervision.
Our model starts with two separate pathways: an appearance pathway that outputs feature-based region segmentation for a single image, and a motion pathway that outputs motion features for a pair of images. It then binds them in a conjoint representation called segment flow that pools flow offsets over each region and provides a gross characterization of moving regions for the entire scene. By training the model to minimize view synthesis errors based on segment flow, our appearance and motion pathways learn region segmentation and flow estimation automatically without building them up from low-level edges or optical flows respectively.
Our model demonstrates the surprising emergence of objectness in the appearance pathway, surpassing prior works on zero-shot object segmentation from an image, moving object segmentation from a video with unsupervised test-time adaptation, and semantic image segmentation by supervised fine-tuning. Our work is the first truly end-to-end zero-shot object segmentation from videos. It not only develops generic objectness for segmentation and tracking, but also outperforms prevalent image-based contrastive learning methods without augmentation engineering.
We learn a single-image segmentation network and a dual-frame motion network with an unsupervised image reconstruction loss. We sample two frames,
Qualitative salient object detection results. We directly transfer our pretrained segmentation network to novel images on the DUTS dataset without any finetuning. Surprisingly, we find that the model pretrained on videos to segment moving objects can generalize to detect stationary unmovable objects in a static image, e.g. the statue, the plate, the bench and the tree in the last column.
