Skip to content

Latest commit

 

History

History
174 lines (147 loc) · 6.8 KB

GETTING_STARTED.md

File metadata and controls

174 lines (147 loc) · 6.8 KB
Raw

Getting Started with DVIS

This document provides a brief intro of the usage of DVIS++ and OV-DVIS++.

Please see Getting Started with Detectron2 for full usage.

Training

We provide two scripts, train_net_video.py and train_net_video_ov.py, which are designed to train all the configs provided in DVIS++ and OV-DVIS++ respectively.

To train a model with train_net_video.py/train_net_video_ov.py, first setup the corresponding datasets following datasets/README.md, then download the pre-trained weights from here and put them in the current working directory. Once these are set up, run:

# For close-vocabulary

# train the segmenter of the DVIS++
python train_net_video.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/CTVIS_R50.yaml \
  --resume MODEL.WEIGHTS /path/to/mask2former_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

# Using contrastive learning during the training process of the segmenter requires a significant amount of GPU memory.
# If the GPU memory requirement cannot be met, contrastive learning can be omitted,
# but it may result in a performance decrease.
python train_net_video.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/MinVIS_R50.yaml \
  --resume MODEL.WEIGHTS /path/to/mask2former_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

# if using multi machines
# at machine 1
python train_net_video.py \
  --num-gpus $NUM_GPUS \
  --num-machines=$NUM_MACHINES \
  --machine-rank 1 \
  --dist-url tcp://$IP_MACHINE1:13325
  --config-file /path/to/CTVIS_R50.yaml \
  --resume MODEL.WEIGHTS /path/to/mask2former_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS*$NUM_MACHINES
# at other machines
python train_net_video.py \
  --num-gpus $NUM_GPUS \
  --num-machines=$NUM_MACHINES \
  --machine-rank $MACHINE_RANK \
  --dist-url tcp://$IP_MACHINE1:13325
  --config-file /path/to/CTVIS_R50.yaml \
  --resume MODEL.WEIGHTS /path/to/mask2former_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS*$NUM_MACHINES

# train the DVIS++(online)
python train_net_video.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/DVIS_Plus_Online_R50.yaml \
  --resume MODEL.WEIGHTS /path/to/segmenter_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

# train the DVIS++(offline)
python train_net_video.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/DVIS_Plus_Offline_R50.yaml \
  --resume MODEL.WEIGHTS /path/to/online_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

# For open-vocabulary

# if only use the COCO dataset for training, you can directly load
# the pre-trained weights of FC-CLIP without fine-tuning the segmenter.

# if use both the COCO dataset and video datasets for training
# finetune the segmenter of the OV-DVIS++
python train_net_video_ov.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/open_vocabulary/FC-CLIP_combine_480p.yaml \
  --resume MODEL.WEIGHTS /path/to/fcclip_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

# train the OV-DVIS++(online)
python train_net_video_ov.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/open_vocabulary/DVIS_Online_combine_480p.yaml \
  --resume MODEL.WEIGHTS /path/to/segmenter_pretrained_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

# train the OV-DVIS++(offline)
python train_net_video_ov.py \
  --num-gpus $NUM_GPUS \
  --config-file /path/to/open_vocabulary/DVIS_Offline_combine_480p.yaml \
  --resume MODEL.WEIGHTS /path/to/online_weights.pth \
  SOLVER.IMS_PER_BATCH $NUM_GPUS

Evaluation

Prepare the datasets following datasets/README.md and download trained weights from here. Once these are set up, run: ``` python train_net_video.py
--num-gpus 8
--config-file /path/to/config.yaml
--eval-only MODEL.WEIGHTS /path/to/weight.pth

# For the ytvis and ovis datasets, 
# you need to submit the predicted results.json to the evaluation server to obtain the evaluation results.

# For the VIPSeg dataset
python utils/eval_vpq_vspw.py --submit_dir output/inference \
--truth_dir datasets/VIPSeg/VIPSeg_720P/panomasksRGB \
--pan_gt_json_file datasets/VIPSeg/VIPSeg_720P/panoptic_gt_VIPSeg_val.json

python utils/eval_stq_vspw.py --submit_dir output/inference \
--truth_dir datasets/VIPSeg/VIPSeg_720P/panomasksRGB \
--pan_gt_json_file datasets/VIPSeg/VIPSeg_720P/panoptic_gt_VIPSeg_val.json

# For the VSPW dataset
python utils/eval_miou_vspw.py datasets/VSPW_480p output/inference

python utils/eval_vc_vspw.py datasets/VSPW_480p output/inference
```

Visualization

  1. Pick a trained model and its config file. To start, you can pick from model zoo, for example, configs/ovis/DVIS_Plus_Offline_R50.yaml.
  2. For DVIS++, we provide demo_video/demo_long_video.py to visualize outputs of a trained model. Run it with: 
    python demo_video/demo_long_video.py \
--config-file /path/to/config.yaml \
--input /path/to/images_folder \
--output /path/to/output_folder \  
--opts MODEL.WEIGHTS /path/to/checkpoint_file.pth

# if the video if long (> 300 frames), plese set the 'windows_size'
python demo_video/demo_long_video.py \
--config-file /path/to/config.yaml \
--input /path/to/images_folder \
--output /path/to/output_folder \  
--windows_size 300 \
--opts MODEL.WEIGHTS /path/to/checkpoint_file.pth
    The input is a folder containing video frames saved as images. For example, ytvis_2019/valid/JPEGImages/00f88c4f0a.
  3. For OV-DVIS++, we provide demo_video/open_vocabulary/demo.py to visualize outputs of a trained open-vocabulary model. Run it with: 
    python demo_video/open_vocabulary/demo.py \
--config-file configs/open_vocabulary/DVIS_Offline_combine_480p_demo.yaml \
--input /path/to/images_folder \
--output /path/to/output_folder \  
--merge \
--thing_classes new_cls_1 new_cls_2 ... new_cls_n \
--stuff_classes new_cls_1 new_cls_2 ... new_cls_n \
--opts MODEL.WEIGHTS /path/to/checkpoint_file.pth \

# "--merge" refers to using a merged vocabulary set of training datasets.
# "--thing_classes" refers to the new "thing" vocabularies you want to add additionally.
# "--stuff_classes" refers to the new "stuff" vocabularies you want to add additionally.
# "--clear" refers to clearing the default vocabulary set and using only the additional vocabulary set provided by you.

# a example
python demo_video/open_vocabulary/demo.py \
--config-file configs/open_vocabulary/DVIS_Offline_combine_480p_demo.yaml \
--input /path/to/images_folder \
--output /path/to/output_folder \  
--merge \
--thing_classes  carrot,carrots lantern,lanterns \
--stuff_classes hay \
--opts MODEL.WEIGHTS /path/to/checkpoint_file.pth \