Official implementation for Pretrain-then-Adapt: Uncertainty-Aware Test-Time Adaptation for Text-based Person Search [SIGIR 2026].
This repository contains the SOTA reproduction code for the paper's Pretrain-then-Adapt experiments:
- CLIP/IRRA-based UATTA on CUHK-PEDES, ICFG-PEDES, and RSTPReid.
- X-VLM/CMP-based UATTA on PAB.
Create an Anaconda environment and install the dependencies:
conda create -n uatta python=3.9 -y
conda activate uatta
# Install a PyTorch build matching your CUDA version.
pip install torch torchvision torchaudio
pip install -r requirements.txtPlease follow the upstream dataset/checkpoint pages for access instructions and license terms.
| Resource | Official link | Notes |
|---|---|---|
| UATTA paper | https://arxiv.org/abs/2604.08598 | Our paper page. |
| IRRA/HAM checkpoint | https://github.com/sssaury/HAM | HAM project/checkpoint entry. |
| CLIP tokenizer vocabulary | https://github.com/openai/CLIP/blob/main/clip/bpe_simple_vocab_16e6.txt.gz | Required by IRRA_UATTA/utils/simple_tokenizer.py. |
| CUHK-PEDES | http://xiaotong.me/static/projects/person-search-language/dataset.html | Official project page. |
| ICFG-PEDES | https://github.com/zifyloo/SSAN | Official dataset/code entry. |
| RSTPReid | https://github.com/NjtechCVLab/RSTPReid-Dataset | Official dataset repository. |
| CMP/X-VLM checkpoint | https://github.com/Shuyu-XJTU/CMP | CMP/PAB official repository. |
| PAB dataset | https://github.com/Shuyu-XJTU/CMP | PAB dataset/code/checkpoint entry. |
Arrange the IRRA-side files as follows:
IRRA_UATTA/
├── data/
│ └── bpe_simple_vocab_16e6.txt.gz
├── checkpoints/
│ └── HAM_checkpoint/
│ └── random100w_2HAMcaptions/
│ └── best0.pth
├── input/
│ └── images/
│ ├── CUHK-PEDES/
│ ├── ICFG-PEDES/
│ └── RSTPReid/
Download the CLIP tokenizer vocabulary before running the IRRA-side experiments:
cd /path/to/OpenSource_Release
mkdir -p IRRA_UATTA/data
curl -L https://github.com/openai/CLIP/raw/main/clip/bpe_simple_vocab_16e6.txt.gz \
-o IRRA_UATTA/data/bpe_simple_vocab_16e6.txt.gzArrange the CMP-side files as follows:
CMP_UATTA/
├── checkpoint/
│ ├── 16m_base_model_state_step_199999.th
│ └── bert-base-uncased/
├── data/
│ └── PAB/
│ ├── annotation/
│ └── image/
Run each experiment separately after activating the uatta environment. The configs evaluate exactly the same epoch set as the retained SOTA logs.
cd /path/to/OpenSource_Release/IRRA_UATTA
CUDA_VISIBLE_DEVICES=0 python tta.py \
--config_file tta_configs/ham_cuhk_tta/uatta_config.yaml \
--seed 42 \
--device cuda \
--output_dir outputs/reproduce/cuhk_pedes \
--eval_epochs 0,1,2,4,9,14,19 \
> logs/cuhk_reproduce.log 2>&1cd /path/to/OpenSource_Release/IRRA_UATTA
CUDA_VISIBLE_DEVICES=0 python tta.py \
--config_file tta_configs/ham_icfg_tta/uatta_config.yaml \
--seed 42 \
--device cuda \
--output_dir outputs/reproduce/icfg_pedes \
--eval_epochs 0,1,2,4,9 \
> logs/icfg_reproduce.log 2>&1cd /path/to/OpenSource_Release/IRRA_UATTA
CUDA_VISIBLE_DEVICES=0 python tta.py \
--config_file tta_configs/ham_rstp_tta/uatta_config.yaml \
--seed 42 \
--device cuda \
--output_dir outputs/reproduce/rstpreid \
--eval_epochs 0,1,2,4,9,14,19,29,39,49,59 \
> logs/rstp_reproduce.log 2>&1cd /path/to/OpenSource_Release/CMP_UATTA
mkdir -p logs outputs/reproduce/pab
CUDA_VISIBLE_DEVICES=0 python tta.py \
--config configs/uatta_config.yaml \
--task pab_reproduce \
--output_dir outputs/reproduce/pab \
--checkpoint checkpoint/16m_base_model_state_step_199999.th \
--seed 42 \
--device cuda \
--eval_epochs 0,1,2,4,9,14,19,29,39,49,59 \
--tta \
> logs/pab_reproduce.log 2>&1The original paper/retained SOTA metrics and the latest reproduced metrics are summarized below. Full adaptation traces are recorded in SOTA_Reproduction_Report.md.
| Dataset | Source | Epoch | Rank@1 | Rank@5 | Rank@10 | mAP | mINP |
|---|---|---|---|---|---|---|---|
| CUHK-PEDES | Paper/retained SOTA | 9 | 70.923 | 86.891 | 91.862 | 63.503 | 47.209 |
| CUHK-PEDES | Latest reproduction | 9 | 70.809 | 86.907 | 91.862 | 63.469 | 47.192 |
| ICFG-PEDES | Paper/retained SOTA | 9 | 62.152 | 77.318 | 82.950 | 36.110 | 5.934 |
| ICFG-PEDES | Latest reproduction | 9 | 62.006 | 77.232 | 82.784 | 35.923 | 5.821 |
| RSTPReid | Paper/retained SOTA | 49 | 61.850 | 81.050 | 88.400 | 46.304 | 22.276 |
| RSTPReid | Latest reproduction | 49 | 61.950 | 81.200 | 88.300 | 46.274 | 22.157 |
| PAB | Paper/retained SOTA | 49 | 76.138 | 98.028 | 99.090 | 86.145 | 86.145 |
| PAB | Latest reproduction | 49 | 75.784 | 98.028 | 98.989 | 86.001 | 86.001 |
Small numeric differences are expected across hardware, CUDA, and PyTorch versions.
This release directly builds on code from:
See THIRD_PARTY_NOTICES.md for direct fork attribution.
We also thank the following repositories for methodological or implementation references:
- 2025-ICLR-TCR: https://github.com/XLearning-SCU/2025-ICLR-TCR
- TENT: https://github.com/DequanWang/tent
- MLLM4Text-ReID: https://github.com/WentaoTan/MLLM4Text-ReID
- HAM: https://github.com/sssaury/HAM
@conference{zhang2026pretrain,
title={Pretrain-then-Adapt: Uncertainty-Aware Test-Time Adaptation for Text-based Person Search},
author={Jiahao Zhang and Shaofei Huang and Yaxiong Wang and Zhedong Zheng},
year={2026},
booktitle={SIGIR},
doi={https://doi.org/10.1145/3805712.3809598},
url={https://arxiv.org/abs/2604.08598},
}For questions, please contact Jiahao Zhang at yc57963@um.edu.mo.