Initialisation takes a long time compared with ORBSLAM2

[martinakos]

I'm running stella_vslam with perspective monocular camera configuration on my own dataset.
ORBSLAM2 can initialise on this dataset very quick.
However, stella_vslam takes a very long time to initialise, if it ever initialises, on the same dataset with same calibration as I use in ORBSLAM2 and same parameters for feature extraction.

The debug logs for stella_vslam keep repeating these sort of messages:

[2024-02-06 11:27:52.083] [D] try to initialize with the initial frame and the current frame: frame 457 - frame 459
[2024-02-06 11:27:52.087] [D] reconstruct_with_F
[2024-02-06 11:27:52.120] [D] try to initialize with the initial frame and the current frame: frame 457 - frame 460
[2024-02-06 11:27:52.123] [D] reconstruct_with_F
[2024-02-06 11:27:52.139] [D] DESTRUCT: initialize::perspective
[2024-02-06 11:27:52.182] [D] CONSTRUCT: initialize::perspective
[2024-02-06 11:27:52.223] [D] try to initialize with the initial frame and the current frame: frame 462 - frame 463
[2024-02-06 11:27:52.226] [D] reconstruct_with_F
[2024-02-06 11:27:52.239] [D] try to initialize with the initial frame and the current frame: frame 462 - frame 464
[2024-02-06 11:27:52.242] [D] reconstruct_with_F
[2024-02-06 11:27:52.279] [D] DESTRUCT: initialize::perspective
[2024-02-06 11:27:52.320] [D] CONSTRUCT: initialize::perspective
[2024-02-06 11:27:52.340] [D] try to initialize with the initial frame and the current frame: frame 466 - frame 467
[2024-02-06 11:27:52.343] [D] reconstruct_with_F
[2024-02-06 11:27:52.384] [D] try to initialize with the initial frame and the current frame: frame 466 - frame 468
[2024-02-06 11:27:52.388] [D] reconstruct_with_F

Eventually at some point the initialization succeeds with F.

I wonder if there is any configuration I can change to make stella_vslam initialise as quickly as ORBSLAM2?

Also, in ORBSLAM2 I see the initialisation features being tracked before the first map is created. I can't see that in stella_vslam so I don't know if it's tracking any features at initialisation. Maybe adding that visualisation would be a nice feature for debugging what's happening during initialisation.

This is my configuration file:

Camera:
  name: "D435i monocular"
  setup: "monocular"
  model: "perspective"

  fx: 424.8605651855469
  fy: 421.48828125
  cx: 424.8605651855469
  cy: 239.51873779296875

  k1: 0.0
  k2: 0.0
  p1: 0.0
  p2: 0.0
  k3: 0.0

  fps: 30.0
  cols: 848
  rows: 480

  color_order: "RGB"

Feature:
  name: "default ORB feature extraction setting"
  max_num_keypoints: 1000
  scale_factor: 1.2
  num_levels: 8
  ini_fast_threshold: 20
  min_fast_threshold: 7

Mapping:
  baseline_dist_thr_ratio: 0.02
  redundant_obs_ratio_thr: 0.9
  num_covisibilities_for_landmark_generation: 20
  num_covisibilities_for_landmark_fusion: 20

and I use the default vocabulary file provided in https://github.com/stella-cv/FBoW_orb_vocab/raw/main/orb_vocab.fbow

[martinakos]

Any help with this issue?

[martinakos]

Maybe related #348

[Skwangles]

I would suggest using an example .yaml config from some of the examples and see if any of those work better.
I'm not a developer for Stella, so I'm not sure re actual implementation reason it might happen.

e.g. https://github.com/stella-cv/stella_vslam/blob/main/example/euroc/EuRoC_mono.yaml