Skip to content

ViSeLA v1.0.0 Models & Data Annotations

Latest

Choose a tag to compare

@ziangbuchu ziangbuchu released this 11 Nov 04:33
· 2 commits to main since this release

ViSeLA v1.0.0 provides the official implementation, four data annotation archives, and two model checkpoints for multilingual text-based person retrieval.

Dataset Download

Download Contents
visela_downstream_multilingual_annos.tar.gz 32 train/test JSON files for CUHK-PEDES, ICFG-PEDES, RSTPReid, and UFineBench in English, Chinese, French, and German; extracts to DATASET/annos/.
visela_downstream_codeswitch_annos.tar.gz 16 code-switch training JSON files for the same four datasets and languages; extracts to DATASET/code_switch/.
multipedes_10m_aligned_parquet.tar.gz 297 aligned HAM and WebPerson Parquet files, plus a portable WebPerson manifest.
multipedes_ham1m_4langs_json.tar.gz HAM four-language JSON with 999,007 images and 1,997,789 caption values per language.

Images are obtained separately from the original datasets under their respective terms. All packaged image paths are relative; each archive includes a layout guide. The downstream filename code ch denotes Chinese.

The aligned Parquet archive contains 1,187,383 HAM and 1,312,617 WebPerson rows per target language, totaling 10,000,000 image-text pairs when counting English once (source_caption) plus German, French, and Chinese. This archive and the separate HAM JSON are released annotation sources; the paper describes an 8M pretraining corpus.

UFineBench code-switch annotations cover all English and Chinese training records; the French and German files contain 45 and 314 fewer records, respectively, than their multilingual counterparts. The HAM JSON contains 652 empty English caption values.

Released Checkpoints

The following checkpoints include their training configurations. Use the embedded configuration for each model and adapt its relative paths to your installation.

Download Training configuration Checkpoint contents
visela_pretrain_ham_luperson_refiner_ep1.th HAM multilingual shards and LUPerson bilingual data, epoch 1; four experts, four queries, refiner.sim_type: lsp_max, routing_temp: 1.0, batch size 72 per GPU. model and config; pretraining weights for downstream fine-tuning.
visela_cuhk_pedes_baseline_router.pth CUHK-PEDES multilingual baseline initialized from CCLM-X2VLM; 10 epochs, four experts, four queries, refiner.sim_type: avg, routing_temp: 0.85, batch size 64 per GPU, text length 77, code-switch probability 0.3. state_dict and config; optimizer state excluded.

Both checkpoints retain the original FP32 model tensors and int64 buffers. The released CUHK-PEDES baseline router has the following retrieval scores from its original training log (percent):

en R@1 ch R@1 fr R@1 de R@1 Avg R@1 Avg mAP
78.22 76.07 76.74 76.10 76.78 68.67

The paper results report ViSeLA with MultiPedes pretraining.

SHA256 Verification

Download SHA256SUMS.txt alongside all four data archives and both model files, then verify them with:

sha256sum -c SHA256SUMS.txt

SHA256SUMS_data.txt provides checksums for the four data archives.

SHA256 values for the six data and model files:

29ac2f58510a242dc1658c91cb1c86eac7719ca957b8705717b48030488657f5  visela_downstream_multilingual_annos.tar.gz
ed126c4075ba6d839e1f24a98e882e2e8ccbd5b61b0655f48be300e25db9f326  visela_downstream_codeswitch_annos.tar.gz
98763463784bce08854a752ee746c0b1b3b80451e271013126b95da87168652b  multipedes_10m_aligned_parquet.tar.gz
8e4f50796205b75163859b8c4b3ae8d10c7ea5220db586a001cb5684c51e6674  multipedes_ham1m_4langs_json.tar.gz
113b687c895fc570eaf773ee0fcadb7321b2bd26fbddc429ed829e3a5f8dfa82  visela_pretrain_ham_luperson_refiner_ep1.th
3776aea5fe499d8c0365173fb73e662ae56edc2314068053f977f1506d22a079  visela_cuhk_pedes_baseline_router.pth

SHA256 values for the checksum files:

23c36680ca824a32776e5999c7013fadc2c3090309dacb8514027a5c13a67dcd  SHA256SUMS.txt
3cbea835620a4bdf4344bb7dcf061af9eb14119f7b64908586277de3cfc9271f  SHA256SUMS_data.txt