astronolan/gz3d-aion-replication-attempt

GalaxyZoo 3D Segmentation Benchmark

An attempted replication of the GZ3D segmentation dataset used in section 7.2.3 in the AION-1. This dataset contains volunteer segmentations for the following channels: center, star, spiral, bar. It also contains the RGB images from Legacy Survey and the tokens from the AION-1 image tokenizer of the Legacy Survey image bands. Notably, these are pre–AION-1 transformer encoder, but post image-codec encoder/quantization. The reason why it's an 'attempted' replication is that we cannot gurantee that the train/val/test split is equivalent to the one used in the paper. Furthermore, this replication was performed autonomously by Codex and the code has not been verified as faithful.

Contents

Each row contains the GZ3D/MaNGA identity and coordinates, the nearest Legacy Survey DR10 South object within 1 arcsec and its separation, GZ3D vote counts, AION-1-B image token IDs (int32[24,24]), and four resampled-for-angular-scale GZ3D segmentation channels (float32[4,96,96]) in center, star, spiral, bar order. rgb (uint8[96,96,3]) is included. It is cut from the exact DR10 South brick JPEG used by the Multimodal Universe builder: the JPEG is flipped into FITS-array orientation, centered at the matched object's g-band WCS pixel, cut to 160x160, then center-cropped to 96x96.

Image tokens use polymathic-ai/aion-base revision 40541618104bab0fa85c8af68daeb867a720bb8c. Coordinate matches use the hugging-science/mmu_legacysurvey_dr10_south_21 HATS catalog at revision 982f321784d8daf2a5f983eb3ec560f31b90d667, thanks HATS/LSDB for making this easy!

Data sources and sample selection

The public Galaxy Zoo 3D v4.0.0 metadata used to construct the sample is available directly from SDSS:

• gz3d_metadata.fits

Rows were selected from that metadata by requiring:

gz_total_classifications >= 20
and
(gz_bar_votes / gz_total_classifications >= 0.2
 or gz_spiral_votes / gz_total_classifications >= 0.2)

The filtered galaxies were cross-matched against Legacy Survey DR10 South. At most the single nearest object within 1.0 arcsecond was retained.

Column descriptions

All columns are non-nullable. Array dimensions and channel order are fixed.

Column	Arrow type / shape	Description
manga_id	string	MaNGA target identifier from the public GZ3D metadata.
gz3d_filename	string	Filename of the corresponding public per-object GZ3D FITS product. The remote file adds a gz3d_ prefix and .gz suffix.
source_row	int64	Zero-based row number in the original public gz3d_metadata.fits, before filtering or sorting.
candidate_order	int64	Zero-based position after the vote filter and deterministic sorting by nested nside-16 HEALPix, then gz3d_filename, then source_row. This sequence covers all 7,614 filtered GZ3D candidates, so it can contain gaps after unmatched candidates are removed.
manifest_order	int64	Contiguous zero-based position after cross-matching and removing candidates without a Legacy Survey match within 1 arcsecond. This is the final dataset order used to assign splits and write shards.
gz3d_ra	float64	GZ3D target right ascension in ICRS degrees.
gz3d_dec	float64	GZ3D target declination in ICRS degrees.
legacy_object_id	string	Identifier of the nearest retained Legacy Survey DR10 South object, in BRICK-OBJID form.
legacy_ra	float64	Right ascension of the matched Legacy Survey object in ICRS degrees.
legacy_dec	float64	Declination of the matched Legacy Survey object in ICRS degrees.
match_sep_arcsec	float64	Great-circle angular separation between the GZ3D and Legacy Survey coordinates, in arcseconds. Values are at most 1.0.
gz_total_classifications	int32	Total catalog morphology classifications for the galaxy. This is a catalog routing/selection count, not the number of GZ3D drawing records or the maximum segmentation value.
gz_bar_votes	int32	Catalog classifications voting for a bar. This routing/selection count is distinct from the pixel-overlap maximum in gz3d_bar_overlap_max_votes.
gz_spiral_votes	int32	Catalog classifications voting for spiral structure. This routing/selection count is distinct from the pixel-overlap maximum in gz3d_spiral_overlap_max_votes.
gz3d_center_overlap_max_votes	int32	Maximum pixel value in the center channel, gz3d_segmentation[0]. This is the largest center-point cluster overlap count anywhere in the processed 96x96 map.
gz3d_star_overlap_max_votes	int32	Maximum pixel value in the foreground-star channel, gz3d_segmentation[1]. This is the largest star-point cluster overlap count anywhere in the processed 96x96 map.
gz3d_spiral_overlap_max_votes	int32	Maximum pixel value in the spiral channel, gz3d_segmentation[2]. This is the largest spiral-drawing overlap count anywhere in the processed 96x96 map.
gz3d_bar_overlap_max_votes	int32	Maximum pixel value in the bar channel, gz3d_segmentation[3]. This is the largest bar-drawing overlap count anywhere in the processed 96x96 map.
tok_image	int32 [24,24]	AION-1-B image-token grid. It contains 576 row-major token IDs generated from a matched 160x160 float32 Legacy Survey g,r,i,z flux cutout. Valid IDs lie in [0, 4374].
gz3d_segmentation	float32 [4,96,96]	GZ3D overlap-count maps in channel order center, star, spiral, bar. Processing and value semantics are described below.
rgb	uint8 [96,96,3]	Aligned Legacy Survey DR10 South RGB visualization in standard red, green, blue channel order. Processing is described below.
split	string	Explicit positional split label: train, threshold_validation, or test.

Ordering fields

The three order columns are retained for auditability:

• source_row connects a row back to the unfiltered public metadata.
• candidate_order records the deterministic order before cross-match failures are removed.
• manifest_order records the final contiguous order and is the position used for split assignment.

Segmentation values

gz3d_segmentation contains volunteer-overlap count maps, not probabilities and not the catalog values in gz_bar_votes or gz_spiral_votes:

• center and star contain filled cluster ellipses whose values reflect the number of clustered volunteer points.
• spiral and bar contain the number of submitted drawing polygons that overlap each pixel.

The public 525x525 maps were cast to uint8, bilinearly resampled from 0.099 arcsecond/pixel to 0.262 arcsecond/pixel, center-cropped to 96x96, vertically flipped, and stored as float32. The stored values remain integer-valued overlap counts. The four gz3d_*_overlap_max_votes columns store the per-row maximum of channels 0 through 3 in the same center, star, spiral, bar order.

RGB construction

The rgb field was not rendered locally from the four flux arrays. It was constructed from the official pre-rendered Legacy Survey DR10 South brick JPEG named legacysurvey-<brick>-image.jpg:

1. Parse the DR10 brick from legacy_object_id.
2. Read the g-band coadd FITS header and transform legacy_ra, legacy_dec to brick pixel coordinates using its WCS.
3. Download and decode the brick JPEG as 8-bit RGB.
4. Vertically flip the JPEG into FITS-array orientation.
5. Extract a strict 160x160 crop centered on the matched object.
6. Center-crop that image to 96x96 and store it as uint8[96,96,3].

Splits

• train: 2,839 rows
• threshold_validation: 142 rows
• test: 568 rows

The split value is also stored in every Parquet row. Split labels are 80/4/16 benchmark split. Not necessarily a perfect replication of the original AION-1 methodology.

GZ3D Citation

@article{Masters_2021,
   title={Galaxy Zoo: 3D – crowdsourced bar, spiral, and foreground star masks for MaNGA target galaxies},
   volume={507},
   ISSN={1365-2966},
   url={http://dx.doi.org/10.1093/mnras/stab2282},
   DOI={10.1093/mnras/stab2282},
   number={3},
   journal={Monthly Notices of the Royal Astronomical Society},
   publisher={Oxford University Press (OUP)},
   author={Masters, Karen L and Krawczyk, Coleman and Shamsi, Shoaib and Todd, Alexander and Finnegan, Daniel and Bershady, Matthew and Bundy, Kevin and Cherinka, Brian and Fraser-McKelvie, Amelia and Krishnarao, Dhanesh and Kruk, Sandor and Lane, Richard R and Law, David and Lintott, Chris and Merrifield, Michael and Simmons, Brooke and Weijmans, Anne-Marie and Yan, Renbin},
   year={2021},
   month=Aug,
   pages={3923–3935}
}

Legacy Dataset Citation

Data Release 10 (DR10) is the tenth public data release of the Legacy Surveys.

When using data from the Legacy Surveys in papers, please use the following acknowledgment:

The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; Proposal ID #2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Prop. ID #2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the Mayall z-band Legacy Survey (MzLS; Prop. ID #2016A-0453; PI: Arjun Dey). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF’s NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. Pipeline processing and analyses of the data were supported by NOIRLab and the Lawrence Berkeley National Laboratory (LBNL). The Legacy Surveys project is honored to be permitted to conduct astronomical research on Iolkam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation.

NOIRLab is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. LBNL is managed by the Regents of the University of California under contract to the U.S. Department of Energy.

This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A&M University, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo a Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cientifico e Tecnologico and the Ministerio da Ciencia, Tecnologia e Inovacao, the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energeticas, Medioambientales y Tecnologicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgenossische Technische Hochschule (ETH) Zurich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ciencies de l’Espai (IEEC/CSIC), the Institut de Fisica d’Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig Maximilians Universitat Munchen and the associated Excellence Cluster Universe, the University of Michigan, NSF’s NOIRLab, the University of Nottingham, the Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A&M University.

BASS is a key project of the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories of China, the Chinese Academy of Sciences (the Strategic Priority Research Program “The Emergence of Cosmological Structures” Grant # XDB09000000), and the Special Fund for Astronomy from the Ministry of Finance. The BASS is also supported by the External Cooperation Program of Chinese Academy of Sciences (Grant # 114A11KYSB20160057), and Chinese National Natural Science Foundation (Grant # 12120101003, # 11433005).

The Legacy Survey team makes use of data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), which is a project of the Jet Propulsion Laboratory/California Institute of Technology. NEOWISE is funded by the National Aeronautics and Space Administration.

The Legacy Surveys imaging of the DESI footprint is supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH1123, by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract; and by the U.S. National Science Foundation, Division of Astronomical Sciences under Contract No. AST-0950945 to NOAO.

CITATION

@ARTICLE{2019AJ....157..168D,
       author = {{Dey}, Arjun and {Schlegel}, David J. and {Lang}, Dustin and {Blum}, Robert and {Burleigh}, Kaylan and {Fan}, Xiaohui and {Findlay}, Joseph R. and {Finkbeiner}, Doug and {Herrera}, David and {Juneau}, St{\'e}phanie and {Landriau}, Martin and {Levi}, Michael and {McGreer}, Ian and {Meisner}, Aaron and {Myers}, Adam D. and {Moustakas}, John and {Nugent}, Peter and {Patej}, Anna and {Schlafly}, Edward F. and {Walker}, Alistair R. and {Valdes}, Francisco and {Weaver}, Benjamin A. and {Y{\`e}che}, Christophe and {Zou}, Hu and {Zhou}, Xu and {Abareshi}, Behzad and {Abbott}, T.~M.~C. and {Abolfathi}, Bela and {Aguilera}, C. and {Alam}, Shadab and {Allen}, Lori and {Alvarez}, A. and {Annis}, James and {Ansarinejad}, Behzad and {Aubert}, Marie and {Beechert}, Jacqueline and {Bell}, Eric F. and {BenZvi}, Segev Y. and {Beutler}, Florian and {Bielby}, Richard M. and {Bolton}, Adam S. and {Brice{\~n}o}, C{\'e}sar and {Buckley-Geer}, Elizabeth J. and {Butler}, Karen and {Calamida}, Annalisa and {Carlberg}, Raymond G. and {Carter}, Paul and {Casas}, Ricard and {Castander}, Francisco J. and {Choi}, Yumi and {Comparat}, Johan and {Cukanovaite}, Elena and {Delubac}, Timoth{\'e}e and {DeVries}, Kaitlin and {Dey}, Sharmila and {Dhungana}, Govinda and {Dickinson}, Mark and {Ding}, Zhejie and {Donaldson}, John B. and {Duan}, Yutong and {Duckworth}, Christopher J. and {Eftekharzadeh}, Sarah and {Eisenstein}, Daniel J. and {Etourneau}, Thomas and {Fagrelius}, Parker A. and {Farihi}, Jay and {Fitzpatrick}, Mike and {Font-Ribera}, Andreu and {Fulmer}, Leah and {G{\"a}nsicke}, Boris T. and {Gaztanaga}, Enrique and {George}, Koshy and {Gerdes}, David W. and {Gontcho}, Satya Gontcho A. and {Gorgoni}, Claudio and {Green}, Gregory and {Guy}, Julien and {Harmer}, Diane and {Hernandez}, M. and {Honscheid}, Klaus and {Huang}, Lijuan Wendy and {James}, David J. and {Jannuzi}, Buell T. and {Jiang}, Linhua and {Joyce}, Richard and {Karcher}, Armin and {Karkar}, Sonia and {Kehoe}, Robert and {Kneib}, Jean-Paul and {Kueter-Young}, Andrea and {Lan}, Ting-Wen and {Lauer}, Tod R. and {Le Guillou}, Laurent and {Le Van Suu}, Auguste and {Lee}, Jae Hyeon and {Lesser}, Michael and {Perreault Levasseur}, Laurence and {Li}, Ting S. and {Mann}, Justin L. and {Marshall}, Robert and {Mart{\'\i}nez-V{\'a}zquez}, C.~E. and {Martini}, Paul and {du Mas des Bourboux}, H{\'e}lion and {McManus}, Sean and {Meier}, Tobias Gabriel and {M{\'e}nard}, Brice and {Metcalfe}, Nigel and {Mu{\~n}oz-Guti{\'e}rrez}, Andrea and {Najita}, Joan and {Napier}, Kevin and {Narayan}, Gautham and {Newman}, Jeffrey A. and {Nie}, Jundan and {Nord}, Brian and {Norman}, Dara J. and {Olsen}, Knut A.~G. and {Paat}, Anthony and {Palanque-Delabrouille}, Nathalie and {Peng}, Xiyan and {Poppett}, Claire L. and {Poremba}, Megan R. and {Prakash}, Abhishek and {Rabinowitz}, David and {Raichoor}, Anand and {Rezaie}, Mehdi and {Robertson}, A.~N. and {Roe}, Natalie A. and {Ross}, Ashley J. and {Ross}, Nicholas P. and {Rudnick}, Gregory and {Safonova}, Sasha and {Saha}, Abhijit and {S{\'a}nchez}, F. Javier and {Savary}, Elodie and {Schweiker}, Heidi and {Scott}, Adam and {Seo}, Hee-Jong and {Shan}, Huanyuan and {Silva}, David R. and {Slepian}, Zachary and {Soto}, Christian and {Sprayberry}, David and {Staten}, Ryan and {Stillman}, Coley M. and {Stupak}, Robert J. and {Summers}, David L. and {Sien Tie}, Suk and {Tirado}, H. and {Vargas-Maga{\~n}a}, Mariana and {Vivas}, A. Katherina and {Wechsler}, Risa H. and {Williams}, Doug and {Yang}, Jinyi and {Yang}, Qian and {Yapici}, Tolga and {Zaritsky}, Dennis and {Zenteno}, A. and {Zhang}, Kai and {Zhang}, Tianmeng and {Zhou}, Rongpu and {Zhou}, Zhimin},
        title = "{Overview of the DESI Legacy Imaging Surveys}",
      journal = {\aj},
     keywords = {catalogs, surveys, Astrophysics - Instrumentation and Methods for Astrophysics},
         year = 2019,
        month = may,
       volume = {157},
       number = {5},
          eid = {168},
        pages = {168},
          doi = {10.3847/1538-3881/ab089d},
archivePrefix = {arXiv},
       eprint = {1804.08657},
 primaryClass = {astro-ph.IM},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2019AJ....157..168D},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}