vos_eval_root.yaml

defaults:
  - model: sam_pt
  - logging: vos_eval
  - _self_

evaluator:
  _target_: sam_pt.vos_eval.evaluator.SamPtEvaluator
  _recursive_: false

dataset: D17 # D16/D17/Y18/Y19/LV1/LV3/MOSE/BDD100K/G
split: val # val/test
simulate_interactive_point_correction: false
masks_batch_size: 100
seed: 72

d16_path: ${hydra:runtime.cwd}/data/DAVIS/2016
d17_path: ${hydra:runtime.cwd}/data/DAVIS/2017
y18_path: ${hydra:runtime.cwd}/data/YouTube2018
y19_path: ${hydra:runtime.cwd}/data/YouTube
lv_path: ${hydra:runtime.cwd}/data/long_video_set
mose_path: ${hydra:runtime.cwd}/data/mose
bdd100k_path: ${hydra:runtime.cwd}/data/bdd100k/vos
generic_path: null # For generic (G) evaluation, point to a folder that contains "JPEGImages" and "Annotations"

input_only_one_gt_mask_point: false # If true, only one gt mask point will be used for evaluation

size: -1 # Resize the shorter side to this size. -1 to use original resolution
longest_size: 1024 # Resize the longest side to this size. null to use original resolution. Must be used with size=-1

flip: false

output: eval_${dataset}_${split} # Path to save the results. If None, will save to the default path
save_all: false # Save all frames. Useful only in YouTubeVOS/long-time video
save_scores: false
save_overlapping_masks: false # Save overlapping masks along with non-overlapping multi-object masks

visualize_results: true # Whether to visualize the results using wandb
verbose_visualisations: false # Whether to visualize the results in a verbose way (e.g. with input GIFs), slower
vid_ids: null # Evaluate only on the videos specified in the list, e.g. [0,1,2] (or vid_ids=\[0,1,2\] in command line)
max_videos_to_visualize: 30 # Max number of videos to visualize, used when visualize_results flag is set, videos with id >= max_videos_to_visualize will not be visualized
vid_ids_to_visualize: [ 0, 1, 2, 15 ] # Videos to visualize, used when visualize_results flag is set, null for all videos
log_fmt: gif # gif/mp4

max_videos: null # Max number of videos to process, useful with the visualize_results flag and for debugging
max_frames: null # Max number of frames to process per video. Useful for debugging

logging:
  exp_id_verbose: ${logging.exp_id}_${dataset}_${split}_${seed}_${now:%Y.%m.%d_%H.%M.%S}


hydra:
  job:
    chdir: True
  run:
    dir: outputs/${logging.exp_id_verbose}