Downloads · 30 days
0
llama-farm/llama-thunderdome-clip-aerial-vit-b16-v4-unified
llama-thunderdome-clip-aerial-vit-b16-v4-unified is a machine learning model from llama-farm. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Fine-tuned OpenAI CLIP ViT-B/16 for aerial-vehicle + person retrieval. v4 unifies all available training sources — UAVDT, VisDrone, AU-AIR (both legacy crops + freshly-extracted 14K crops), VEDAI, and the drone-fleet…
Downloads · 30 days
0
Access
Public
Updated May 27, 2026
Repo size
2 GB
Likes
0
Public
Click a slice to open those files.
.pt1.2 GB · 59%
From the Hugging Face model README
Fine-tuned OpenAI CLIP ViT-B/16 for aerial-vehicle + person retrieval. v4 unifies all available training sources — UAVDT, VisDrone, AU-AIR (both legacy crops + freshly-extracted 14K crops), VEDAI, and the drone-fleet batch from the 2026-05-25 Drive uploads.
| Model | Top-1 | Top-5 | Best on |
|---|---|---|---|
| v4 (this) | 74.8% | 99.9% | every source |
| v3 (drone-fleet only) | 50.4% | 97.6% | drone-batch (89%) |
| v2 (UAVDT/VisDrone/AU-AIR/VEDAI) | 60.0% | 92.4% | bus (96%) |
| Source | n | top-1 |
|---|---|---|
| vedai | 21 | 0.905 |
| drone-batch | 178 | 0.854 |
| au-air | 617 | 0.776 |
| visdrone | 871 | 0.730 |
| uavdt | 313 | 0.668 |
| Class | n | top-1 |
|---|---|---|
| motorcycle | 11 | 1.000 |
| bus | 222 | 0.982 |
| van | 98 | 0.898 |
| person | 226 | 0.889 |
| bicycle | 33 | 0.879 |
| truck | 494 | 0.846 |
| car | 916 | 0.579 |
ViT-B-16 (openai pretrained)| File | Size | Purpose |
|---|---|---|
best.pt / last.pt | 598 MB | PyTorch fine-tuned weights |
clip-aerial-vit-b16-v4-224.onnx | 345 MB | Clean ONNX (x86/ARM CPU/GPU) |
clip-aerial-vit-b16-v4-surgical.onnx | 344 MB | Hailo-DFC-compatible ONNX |
clip-aerial-vit-b16-v4-224-hailo10h.hef | 75.8 MB | Compiled HEF for Hailo-10H |
eval_results.json | — | Full per-source / per-class eval breakdown |
training_meta.json | — | Hyperparameter snapshot |
results.csv | — | Per-epoch train/val loss + recall |
Same surgical-ONNX + DFC-compile path as v3:
thunderdome clip onnx-surgery --our-pt best.pt --output v4-surgical.onnx ...
thunderdome clip compile-hef --onnx v4-surgical.onnx ... --target hailo10h
thunderdome clip validate-hef-vit --quant-har v4_quantized.har --reference-pt best.pt ...
| Metric | v3 | v4 (this) |
|---|---|---|
| Cosine mean | 0.9551 | 0.8577 |
| Cosine min/max | 0.93 / 0.97 | 0.79 / 0.94 |
| Overlap@5 (ranking) | 0.820 | 0.790 |
| Verdict | PASS | PASS |
V4 sits a touch below v3 on PT-vs-HEF cosine because the 11×-bigger training corpus drifted the weights further from the OpenAI baseline (Hailo's reference ONNX was built against that baseline). Retrieval rankings are still preserved (overlap@5 = 0.79), which is what matters downstream. Compile parameters were identical: Hailo MZ alls config, calibration on 500 crops from the unified pool.
The HEF has the OpenAI CLIP normalization baked into the graph via norm1. Pass letterbox-padded raw uint8 input:
from PIL import Image
def letterbox(img, sz=224, fill=(114, 114, 114)):
w, h = img.size
s = sz / max(w, h)
nw, nh = int(round(w*s)), int(round(h*s))
img = img.resize((nw, nh), Image.BILINEAR)
canvas = Image.new("RGB", (sz, sz), fill)
canvas.paste(img, ((sz-nw)//2, (sz-nh)//2))
return canvas
# Pass to HEF as np.uint8 [1, 224, 224, 3] (NHWC) — DO NOT pre-normalize.
from thunderdome.tracking.reid_clip import detect_clip_arch, CLIPInference
arch = detect_clip_arch("best.pt") # "ViT-B-16"
reid = CLIPInference("best.pt", base_model="ViT-B-16", pretrained="openai")
features = reid.get_features(xyxys, frame) # [N, 512]
Wraps automatically through thunderdome.tracking.clip_reid_adapter.CLIPReidAdapter
when used as a boxmot ReID backend (BotSORT, StrongSORT).
clip-aerial-rn50 repo)If you use this model, please cite llama-thunderdome and the underlying
OpenCLIP work.
🤖 Generated with Claude Code.