xxayt / cstm

CSTM

Implementation for Collaborative Spatial-Temporal Modeling for Language-Queried Video Actor Segmentation, CVPR 2021. PDF.

Setup

• Python 3.8
• PyTorch 1.9
• CUDA 11

Download A2D Sentences dataset from this link. Modify the paths of dataset in ./datasets/a2d_dataset.py. The GloVe word embeddings can be downloaded from this link and put glove.840B.300d.zip in ./word_embedding.

Running

We provide the trained checkpoint of our model in Baidu Drive, password: fs58.

For testing on A2D Sentences dataset, please use the following command:

python test.py \
--data_margin 2 --batch_size 8 \
--gpu_id 0 --resize 320 --skip single \
--dim_semantic 512 \
--dataset A2D --model_root cstm_checkpoints \
--checkpoint checkpoint.pth.tar