Research map / Label-efficient and robust learning / Domain adaptation and generalization

Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic

ICLR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the Alice benchmarks, two real-world target re-ID datasets (AlicePerson, AliceVehicle) paired with editable synthetic sources (PersonX, VehicleX) and an online evaluation server, and benchmarks common synthetic-to-real domain adaptation methods.

Problem

Synthetic-to-real domain adaptation for object re-identification lacks realistic real-world target benchmarks and a fair online evaluation protocol, and existing target sets like Market-1501 have strong clusterability that favors pseudo-label methods.

Approach

Provide real-world target training sets without identity labels and with unassured clusterability, paired with editable synthetic sources, hidden-label test sets, and an online server, then evaluate content-level, pixel-level, feature-level and pseudo-label DA methods on them.

Key results

• MMT reaches 79.46% mAP on PersonX→Market-1501 but only 8.89% mAP on PersonX→AlicePerson, below SPGAN's 10.24%.
• ECN achieves the best AlicePerson result among pseudo-label methods at 14.36% mAP (+8.14% over IDE baseline).
• Attribute descent combined with SPGAN gives +2.39% mAP over SPGAN alone on vehicle re-ID.

Measured effects reported in the paper

Surprising findings

• Pseudo-label methods such as MMT, strong on Market-1501, underperform pixel-level SPGAN on AlicePerson because the target training set lacks clusterability.
• Attribute descent alone gives only small gains (+0.19% mAP on AlicePerson) but larger gains when combined with pixel-level DA.

Limitations

• Benchmark covers only person and vehicle re-ID with two synthetic engines.
• Test labels are hidden, so reproducibility of evaluation depends on the online server.
• Content-level DA gains are modest and FID-based optimization is a limited similarity measure.

Experimental setup

Datasets: PersonX VehicleX Market-1501 VeRi-776 MSMT17

Benchmarks: AlicePerson AliceVehicle Market-1501 VeRi-776

Metrics: mAP rank-1 rank-5 rank-10

Baselines: IDE PCB MSINet PUL ECN UDA MMT CycleGAN SPGAN

Builds on: attribute descent CycleGAN SPGAN MMT PersonX

Closest papers by meaning

• Large-Scale Training Data Search for Object Re-Identification (CVPR 2023 · similarity 0.91)
• Camera-Aware Proxies for Unsupervised Person Re-Identification (AAAI 2021 · similarity 0.88)
• UnrealPerson: An Adaptive Pipeline Towards Costless Person Re-Identification (CVPR 2021 · similarity 0.88)
• Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-Identification (CVPR 2019 · similarity 0.87)
• Real-World Person Re-Identification via Degradation Invariance Learning (CVPR 2020 · similarity 0.87)
• Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.87)