search

SYSU-STAR / H2-Mapping

H2-Mapping: Real-time Dense Mapping Using Hierarchical Hybrid Representation (2023 RAL Best Paper Award)
GNU General Public License v3.0
362 stars 12 forks source link

Why is this system vague in modeling small objects' details? Can parameters be adjusted to enhance clarity? #28

Closed hesen3 closed 3 months ago

hesen3 commented 4 months ago

log_dir: './logs' decoder: parallel_hash_net update_pose: False

criteria: rgb_weight: 0.5 # .5 depth_weight: 1 sdf_weight: 50000.0 fs_weight: 10.0 sdf_truncation: 0.05 # 0.1

decoder_specs: voxel_size: 0.1 # same as mapper_specs L: 4 # Number of levels F_entry: 2 # Number of feature dimensions per entry log2_T: 19 # each level's hashmap_size = F_entry * (2*F_entry) b: 2.0 # each level's resolution = N_min (b**Level)

mapper_specs: start_frame: 0 end_frame: -1 N_rays_each: 4096 # mapping's sampling ray batch_size: 10240 num_vertexes: 200000 inflate_margin_ratio: 0.1 voxel_size: 0.1 step_size: 0.1 num_iterations: 8 max_voxel_hit: 10 final_iter: 0 mesh_res: 8 overlap_th: 0.8 kf_window_size: 8 kf_selection_method: "multiple_max_set_coverage" # "random” or “multiple_max_set_coverage” kf_selection_random_radio: 0.5 # random keyframe ratio insert_method: "intersection" # "naive" or "intersection" insert_ratio: 0.85 offset: 10 # used to make make the coordinate of each point positive use_adaptive_ending: True # adaptive iteration

ros_args: intrinsic: [ 601.347290039062, 601.343017578125, 329.519226074219, 238.586654663086 ] # K[0, 0], K[1, 1], K[0, 2], K[1, 2] color_topic: '/camera/color/image_raw' depth_topic: '/camera/aligned_depth_to_color/image_raw' pose_topic: /vins_estimator/cam_pose

debug_args: verbose: false mesh_freq: -1 render_freq: -1 save_ckpt_freq: -1 render_res: [ 320, 240 ]

JIANG-CX commented 4 months ago

Could you please provide some examples?

hesen3 commented 4 months ago
image image
JIANG-CX commented 4 months ago

In addition to the mapping algorithm, the accuracy of pose estimation and the presence of image motion blur also impact the quality of the reconstruction results.

hesen3 commented 4 months ago

Is this the best result achievable with the "tower_compress.bag" provided by you?

JIANG-CX commented 4 months ago

You can try to increase the “N_rays_each”, “num_iterations” and “mesh_res”. it may get a better result, but it will slow down the mapping process. Or you can modify the tracking module to get the more precise pose.

获取 Outlook for iOShttps://aka.ms/o0ukef

发件人: biosen3 @.> 发送时间: Friday, March 8, 2024 4:56:14 PM 收件人: SYSU-STAR/H2-Mapping @.> 抄送: JIANG Chenxing @.>; Comment @.> 主题: Re: [SYSU-STAR/H2-Mapping] Why is this system vague in modeling small objects' details? Can parameters be adjusted to enhance clarity? (Issue #28)

Is this the best result achievable with the "tower_compress.bag" provided by you?

― Reply to this email directly, view it on GitHubhttps://github.com/SYSU-STAR/H2-Mapping/issues/28#issuecomment-1985301423, or unsubscribehttps://github.com/notifications/unsubscribe-auth/APMKKNV5JCEKJAXCWET2QLDYXF4K5AVCNFSM6AAAAABEMNKIZGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOBVGMYDCNBSGM. You are receiving this because you commented.Message ID: @.***>

hesen3 commented 4 months ago

Do your systems require pre-training?

JIANG-CX commented 4 months ago

No, it is training in the mapping process.

发件人: biosen3 @.> 发送时间: 2024年3月12日 19:28 收件人: SYSU-STAR/H2-Mapping @.> 抄送: JIANG Chenxing @.>; Comment @.> 主题: Re: [SYSU-STAR/H2-Mapping] Why is this system vague in modeling small objects' details? Can parameters be adjusted to enhance clarity? (Issue #28)

Do your systems require pre-training?

― Reply to this email directly, view it on GitHubhttps://github.com/SYSU-STAR/H2-Mapping/issues/28#issuecomment-1991436597, or unsubscribehttps://github.com/notifications/unsubscribe-auth/APMKKNRZVNJKXJD7ZOAI3NDYX3RETAVCNFSM6AAAAABEMNKIZGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOJRGQZTMNJZG4. You are receiving this because you commented.Message ID: @.***>