{"paper":{"title":"Language-driven Semantic Segmentation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LSeg aligns per-pixel image embeddings contrastively with text label embeddings to enable zero-shot semantic segmentation.","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Boyi Li, Kilian Q. Weinberger, Ren\\'e Ranftl, Serge Belongie, Vladlen Koltun","submitted_at":"2022-01-10T18:59:10Z","abstract_excerpt":"We present LSeg, a novel model for language-driven semantic image segmentation. LSeg uses a text encoder to compute embeddings of descriptive input labels (e.g., \"grass\" or \"building\") together with a transformer-based image encoder that computes dense per-pixel embeddings of the input image. The image encoder is trained with a contrastive objective to align pixel embeddings to the text embedding of the corresponding semantic class. The text embeddings provide a flexible label representation in which semantically similar labels map to similar regions in the embedding space (e.g., \"cat\" and \"fu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate that our approach achieves highly competitive zero-shot performance compared to existing zero- and few-shot semantic segmentation methods, and even matches the accuracy of traditional segmentation algorithms when a fixed label set is provided.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the contrastive alignment learned on seen classes will transfer to arbitrary unseen text labels without retraining or additional samples, relying on the semantic structure already present in the pre-trained text encoder.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LSeg achieves competitive zero-shot semantic segmentation by contrastively aligning dense pixel embeddings from a transformer with text embeddings of class labels.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LSeg aligns per-pixel image embeddings contrastively with text label embeddings to enable zero-shot semantic segmentation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"931e15df43fde9c8d345e492bc67159bc6d2959111072035fb79ed708b16bb4e"},"source":{"id":"2201.03546","kind":"arxiv","version":2},"verdict":{"id":"b71a28ff-1f43-47dc-923d-048ee4939a49","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T03:17:04.329514Z","strongest_claim":"We demonstrate that our approach achieves highly competitive zero-shot performance compared to existing zero- and few-shot semantic segmentation methods, and even matches the accuracy of traditional segmentation algorithms when a fixed label set is provided.","one_line_summary":"LSeg achieves competitive zero-shot semantic segmentation by contrastively aligning dense pixel embeddings from a transformer with text embeddings of class labels.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the contrastive alignment learned on seen classes will transfer to arbitrary unseen text labels without retraining or additional samples, relying on the semantic structure already present in the pre-trained text encoder.","pith_extraction_headline":"LSeg aligns per-pixel image embeddings contrastively with text label embeddings to enable zero-shot semantic segmentation."},"references":{"count":13,"sample":[{"doi":"","year":null,"title":"Rethinking Atrous Convolution for Semantic Image Segmentation","work_id":"6f5d4c68-8df6-4794-b125-a10bfe8d5876","ref_index":1,"cited_arxiv_id":"1706.05587","is_internal_anchor":true},{"doi":"","year":2009,"title":"Imagenet: A large-scale hierarchical image database","work_id":"685985d1-7fbd-4ef7-9bd6-a8d79824b9be","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Recent advances in open set recognition: A survey","work_id":"9eb68618-f32a-4c71-a302-d7a65a7fe369","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Open-vocabulary Object Detection via Vision and Language Knowledge Distillation","work_id":"59541f32-18cf-4328-ad60-c9018e1401cf","ref_index":4,"cited_arxiv_id":"2104.13921","is_internal_anchor":true},{"doi":"","year":2013,"title":"Few-shot open-set recognition using meta-learning","work_id":"7a1c13b4-e227-4fa9-8eed-0085512f8fac","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":13,"snapshot_sha256":"7276fb5d6e191c7fa0e4e34fbbbe06f3aacdc7db81c536158414d77ac891d0fd","internal_anchors":4},"formal_canon":{"evidence_count":3,"snapshot_sha256":"f174111fa4f88a18a99f42719ad6b60ecedadcb926f8d1a7455bead88c2639e6"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}