{ "id": 294755, "title": "Quick review: EMBEDSEG", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/294755", "author_name": "John Doe", "post_date": "2021-12-12T13:21:48.729000", "votes": 11, "comment_count": 0, "views": 0, "content": "

I skimmed EMBEDSEG paper and understand the basic concept.
\nHere is summary.

\n

Underlying technology

\n", "messages": [ { "id": 1615626, "postDate": "2021-12-12T13:21:48.730Z", "content": "

I skimmed EMBEDSEG paper and understand the basic concept.
\nHere is summary.

\n

Underlying technology

\n", "rawMarkdown": "I skimmed [EMBEDSEG paper](https://arxiv.org/pdf/2101.10033.pdf) and understand the basic concept.\nHere is summary.\n\n#### Underlying technology\n- [spatial embeddings](https://en.wikipedia.org/wiki/Spatial_embedding)\n*Conceptually spatial embeddings involves a mathematical [embedding](https://en.wikipedia.org/wiki/Embedding) from a space with many dimensions per geographic object to a continuous vector space with a much lower dimension*.\n**EMBEDSEG** is *a variation of the inspiring work in ([7](https://arxiv.org/abs/1906.11109)), a very compact model for end-to-end instance segmentation*.\nActually when I compare the source codes for clustering between 2 methods, EMBEDSEG borrows the algorithm and the source code(Both uses the same function *cluster_with_gt*).\nFrom the paper, I guess *gt* could be ground truth.\n[Related code for EMBEDSEG](https://github.com/juglab/EmbedSeg/blob/50f23233cf9564ff443c67c45a611ce665571c12/EmbedSeg/utils/utils.py#L130) / [Related code for [7]](https://github.com/davyneven/SpatialEmbeddings/blob/39ea2492d43bee59c8f27c06d499f601324c4818/src/utils/utils.py#L105)\n\n\n- [Medoid](https://en.wikipedia.org/wiki/Medoid)\n**EMBEDSEG** uses Medoid instead of [centroid](https://en.wikipedia.org/wiki/Centroid) because centroid *has the unfortunate property that it can lie outside of its defining object* according to this article.\nMaybe this could be core technology.\nThat's because other parts seems to the same as [[7](https://arxiv.org/abs/1906.11109)].\n[Related code for EMBEDSEG](https://github.com/juglab/EmbedSeg/blob/50f23233cf9564ff443c67c45a611ce665571c12/EmbedSeg/utils/generate_crops.py#L84)\n\n\n- [ERF-Net](http://www.robesafe.uah.es/personal/eduardo.romera/pdfs/Romera17tits.pdf)\nEMBEDSEG uses *branched* ERF-Net which is proposed by [[7](https://arxiv.org/abs/1906.11109)].\nThe first branch is for *offsets and the clustering bandwidths*.\nThe second branch is for *‘seediness’ score of the pixel*.\n[Related code for EMBEDSEG](https://github.com/juglab/EmbedSeg/blob/main/EmbedSeg/models/erfnet.py) / [Related code for [7]](https://github.com/davyneven/SpatialEmbeddings/blob/master/src/models/erfnet.py)", "votes": 11 } ], "comments": [], "raw_markdown_by_id": { "1615626": "I skimmed [EMBEDSEG paper](https://arxiv.org/pdf/2101.10033.pdf) and understand the basic concept.\nHere is summary.\n\n#### Underlying technology\n- [spatial embeddings](https://en.wikipedia.org/wiki/Spatial_embedding)\n*Conceptually spatial embeddings involves a mathematical [embedding](https://en.wikipedia.org/wiki/Embedding) from a space with many dimensions per geographic object to a continuous vector space with a much lower dimension*.\n**EMBEDSEG** is *a variation of the inspiring work in ([7](https://arxiv.org/abs/1906.11109)), a very compact model for end-to-end instance segmentation*.\nActually when I compare the source codes for clustering between 2 methods, EMBEDSEG borrows the algorithm and the source code(Both uses the same function *cluster_with_gt*).\nFrom the paper, I guess *gt* could be ground truth.\n[Related code for EMBEDSEG](https://github.com/juglab/EmbedSeg/blob/50f23233cf9564ff443c67c45a611ce665571c12/EmbedSeg/utils/utils.py#L130) / [Related code for [7]](https://github.com/davyneven/SpatialEmbeddings/blob/39ea2492d43bee59c8f27c06d499f601324c4818/src/utils/utils.py#L105)\n\n\n- [Medoid](https://en.wikipedia.org/wiki/Medoid)\n**EMBEDSEG** uses Medoid instead of [centroid](https://en.wikipedia.org/wiki/Centroid) because centroid *has the unfortunate property that it can lie outside of its defining object* according to this article.\nMaybe this could be core technology.\nThat's because other parts seems to the same as [[7](https://arxiv.org/abs/1906.11109)].\n[Related code for EMBEDSEG](https://github.com/juglab/EmbedSeg/blob/50f23233cf9564ff443c67c45a611ce665571c12/EmbedSeg/utils/generate_crops.py#L84)\n\n\n- [ERF-Net](http://www.robesafe.uah.es/personal/eduardo.romera/pdfs/Romera17tits.pdf)\nEMBEDSEG uses *branched* ERF-Net which is proposed by [[7](https://arxiv.org/abs/1906.11109)].\nThe first branch is for *offsets and the clustering bandwidths*.\nThe second branch is for *‘seediness’ score of the pixel*.\n[Related code for EMBEDSEG](https://github.com/juglab/EmbedSeg/blob/main/EmbedSeg/models/erfnet.py) / [Related code for [7]](https://github.com/davyneven/SpatialEmbeddings/blob/master/src/models/erfnet.py)" } }