scout-eccv-itm-v3

scout-eccv-itm-v3

Part of SCOUT (Sim-to-real Contrastive Outlier Understanding via Text), an ECCV 2026 workshop paper on JEPA-based text-based person re-identification for the AI City Challenge Track 4 (Sim2Real). Code: https://github.com/abtraore/SCOUT-ECCV Paper: SCOUT: Sim-to-Real Text-Based Person Retrieval by Embedding-Space Prediction over Frozen Video Features

ScoutITM-v3: a single-stream cross-encoder (ExPLoRA V-JEPA2 + LoRA-adapted EmbeddingGemma token states, bidirectional fusion transformer + ITC/ITM heads) used as a reranker in the paper's headline board system.

• Val mAP@10: 0.8858
• Val Hit@10: 0.9854
• Trainable / total parameters: 120.3M / 656.1M

Blend of ITC-retrieval order and ITM match logit (a=0.5); pure ITM re-ranking alone is worse than ITC alone, always use the blend. See system/README.md.

What's in this repo

Only the parameters SCOUT actually trained: a model.safetensors (trainable predictor weights, and for ExPLoRA/LoRA variants, the adapted encoder deltas) plus a config.json recording the exact recipe. Frozen backbone weights (V-JEPA2 from Meta, the text encoder from its own source) are not included here; they're downloaded from their own public sources at load time.

Usage

from scripts.load_from_hf import load_from_hf

model = load_from_hf("Abdrah/scout-eccv-itm-v3").eval()
# model.encode_image(pixel_values)  # [B, 3, 384, 384] in [0, 1] -> [B, out_dim] L2-normalized
# model.encode_text(["a caption"])  # -> [B, out_dim] L2-normalized

Or drop it straight into the repo's own evaluation script:

uv run python scripts/local_eval.py --hf-repo Abdrah/scout-eccv-itm-v3 --pab-root /path/to/PAB --query-source val

License

Apache 2.0 (see the source repo's LICENSE/NOTICE). This checkpoint does not redistribute any frozen backbone weights, which carry their own upstream licenses.