Bad-RFs icon indicating copy to clipboard operation
Bad-RFs copied to clipboard

colmap运行错误

Open FCInter opened this issue 3 months ago • 1 comments

执行了这个

ns-process-data images --data data/llff_data/blurtanabata/images --output-dir data/bad-nerf/blurtanabata

报错:

[11:22:17] 🎉 Done copying images with prefix 'frame_'.                                        process_data_utils.py:336
[11:22:18] 🎉 Done extracting COLMAP features.                                                       colmap_utils.py:137
──────────────────────────────────────────────  💀 💀 💀 ERROR 💀 💀 💀  ───────────────────────────────────────────────
Error running command: colmap vocab_tree_matcher --database_path data/bad-nerf/blurtanabata/colmap/database.db 
--SiftMatching.use_gpu 1 --VocabTreeMatching.vocab_tree_path 
"/home/mypath/.local/share/nerfstudio/vocab_tree.fbow"
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
I0314 11:22:19.168327 106309 misc.cc:198] 
==============================================================================
Vocabulary tree feature matching
==============================================================================
E0314 11:22:19.170003 106310 logging.h:125]  Check failed: max_num_matches > 0 (0 vs. 0)
E0314 11:22:19.170055 106310 sift.cc:1219] Check failed: options_.Check() 
terminate called after throwing an instance of 'std::invalid_argument'
  what():   Check failed: options_.Check() 
*** Aborted at 1710415339 (unix time) try "date -d @1710415339" if you are using GNU date ***
PC: @     0x7f29f6afc00b gsignal
*** SIGABRT (@0x13c7f00019f36) received by PID 106294 (TID 0x7f29e94fb000) from PID 106294; stack trace: ***
    @     0x7f29f8305631 (unknown)
    @     0x7f29f82da420 (unknown)
    @     0x7f29f6afc00b gsignal
    @     0x7f29f6adb859 abort
    @     0x7f29f6eb3911 (unknown)
    @     0x7f29f6ebf38c (unknown)
    @     0x7f29f6ebf3f7 std::terminate()
    @     0x7f29f6ebf6a9 __cxa_throw
    @     0x562c7cde9d1a colmap::LogMessageFatalThrow<>::~LogMessageFatalThrow()
    @     0x562c7d561606 colmap::(anonymous namespace)::SiftGPUFeatureMatcher::Create()
    @     0x562c7d561701 colmap::CreateSiftFeatureMatcher()
    @     0x562c7ceadeb0 colmap::FeatureMatcherWorker::Run()
    @     0x562c7cf56e91 colmap::Thread::RunFunc()
    @     0x7f29f6eebde4 (unknown)
    @     0x7f29f82ce609 start_thread
    @     0x7f29f6bd8353 clone

我在我自己的图片数据集上跑colmap相同的指令

colmap vocab_tree_matcher --database_path /home/mypath/projects/data/vol01/database.db --SiftMatching.use_gpu 1 --VocabTreeMatching.vocab_tree_path "/home/mypath/.local/share/nerfstudio/vocab_tree.fbow"

这个就没错。求问是什么情况。感谢!

FCInter avatar Mar 14 '24 11:03 FCInter

你好,可能是因为所有图片都是模糊的,导致COLMAP重建失败,这个我们有遇到过,一般多跑几次就能重建出来,如果一直不行,可能是COLMAP的环境有问题,可以尝试nerfstudio官方的docker

另外之前BAD-NeRF的setting是只在模糊图片上跑COLMAP的,这样其实更有挑战性,但是后面我们发现包括Deblur-NeRF的真实数据集里,很多都是直接在模糊+清晰的novel views上面一起跑的COLMAP,这样COLMAP就更容易跑出来,所以我们又重新上传了这个新的Deblur-NeRF合成数据集,是我们用COLMAP在BAD-NeRF模糊+清晰的novel views上面(带GT的相机内参)重新跑过的,也搞了一个新的Data parser,可以直接用,直接在这个合成数据集和Deblur-NeRF的真实数据集上开始训练,就不用自己跑COLMAP了。

LingzheZhao avatar Mar 15 '24 14:03 LingzheZhao