MENSA: Multi-Dataset Harmonized Pretraining for Semantic Segmentation

Existing pretraining methods for semantic segmentation are hampered by the task gap between global image -level pretraining and local pixel-level finetuning. Joint dense-level pretraining is a promising alternative to exploit off-the-shelf annotations from diverse segmentation datasets but suffers from low-quality class embeddings and inconsistent data and supervision signals across multiple datasets by directly employing CLIP. To overcome these challenges, we propose a novel <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</u>ulti-datas<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</u>t harmo<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">N</u>ized pretraining framework for <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</u>emantic s<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</u>gmentation (MENSA). MENSA incorporates high-quality language embeddings and momentum-updated visual embeddings to effectively model the class relationships in the embedding space and thereby provide reliable supervision information for each category. To further adapt to multiple datasets, we achieve one-to-many pixel-embedding pairing with cross-dataset multi-label mapping through cross-modal information exchange to mitigate inconsistent supervision signals and introduce region-level and pixel-level cross-dataset mixing for varying data distribution. Experimental results demonstrate that MENSA is a powerful foundation segmentation model that consistently outperforms popular supervised or unsupervised ImageNet pretrained models for various benchmarks under standard fine-tuning. Furthermore, MENSA is shown to significantly benefit frozen-backbone fine-tuning and zero-shot learning by endowing pixel-level distinctiveness to learned representations.