BIUS-batch1/Keyframes-Qwen3-VL-Embeddings

Keyframes-v13 Multimodal Vector Embeddings (Qwen3-VL-Embedding-2B)

Official 2048-dimensional multimodal vector representations extracted from the BIUS-batch1/Keyframes (v13) dataset using the state-of-the-art Qwen/Qwen3-VL-Embedding-2B vision-language foundation model.

---

📌 Dataset Overview

• Source Keyframes: BIUS-batch1/Keyframes (keyframes-v13.tar.zst)
• Total Keyframes Indexed: 168,050 images
• Total Videos: 873 videos (Batch 1: L21_V001 through L30_V041)
• Model Backbone: Qwen/Qwen3-VL-Embedding-2B
• Embedding Dimension: 2048-d (float32)
• Vector Normalization: L2 Normalized ($|v|_2 = 1.0$)
• Archive Size: ~605 MB (compressed) / ~1.31 GB (uncompressed)

---

📂 Repository Structure

The dataset provides both individual per-video folders and a single compressed archive for fast bulk downloading:

.
├── README.md
├── keyframes_v13_embeddings.tar.zst    # Complete dataset archive (~605 MB)
└── keyframes_embeddings/
    ├── manifest.json                   # Global index of all 873 videos & frame counts
    ├── L21_V001/
    │   ├── L21_V001.npy                # float32 array of shape (N_keyframes, 2048)
    │   └── L21_V001.json               # Frame indices, image IDs, and source paths
    ├── L21_V002/
    │   ├── L21_V002.npy
    │   └── L21_V002.json
    └── ... (873 video folders)

JSON Metadata Schema ({video_id}.json)

{
  "video_id": "L21_V001",
  "num_frames": 490,
  "embedding_dim": 2048,
  "frames": [
    {
      "global_index": 0,
      "frame_idx": 0,
      "image_id": "L21_V001_001_0000000000",
      "path": "/data/aic/shared/artifacts/keyframes-v13/L21_V001/001_0000000000.jpg"
    }
  ]
}

---

🚀 Quickstart & Usage

1. Download Dataset with huggingface_hub

from huggingface_hub import hf_hub_download
import tarfile

# Download the unified archive
archive_file = hf_hub_download(
    repo_id="BIUS-batch1/Keyframes-Qwen3-VL-Embeddings",
    filename="keyframes_v13_embeddings.tar.zst",
    repo_type="dataset"
)
print(f"Downloaded archive to: {archive_file}")

2. Loading Embeddings for a Specific Video

import numpy as np
import json

# Load embedding matrix for video L21_V001
embeddings = np.load("keyframes_embeddings/L21_V001/L21_V001.npy")
print(f"Loaded embeddings shape: {embeddings.shape}")  # (490, 2048)

with open("keyframes_embeddings/L21_V001/L21_V001.json") as f:
    meta = json.load(f)
print(f"Total keyframes: {meta['num_frames']}")

3. Text-to-Image Cross-Modal Retrieval

import numpy as np
from sentence_transformers import SentenceTransformer

# 1. Load query encoder
model = SentenceTransformer("Qwen/Qwen3-VL-Embedding-2B", trust_remote_code=True, device="cuda")

# 2. Encode text query into 2048-d vector
query = "Người dẫn chương trình bản tin thời sự trong trường quay"
query_vec = model.encode([query], normalize_embeddings=True).astype(np.float32)  # shape (1, 2048)

# 3. Compute cosine similarity against video keyframes
video_emb = np.load("keyframes_embeddings/L21_V001/L21_V001.npy")  # shape (N, 2048)
scores = np.dot(video_emb, query_vec.T).squeeze()  # Dot-product equals Cosine Similarity

best_idx = np.argmax(scores)
print(f"Top match in video L21_V001: Frame index {best_idx} with score: {scores[best_idx]:.4f}")

---

��️ Extraction Specifications

Specification	Value
Model	Qwen/Qwen3-VL-Embedding-2B
Input Resolution	$512 \times 288$ RGB Keyframes
Precision	Float16 GPU Inference $\to$ Float32 L2-Normalized Storage
Batch Size	64
Total Runtime	~81.9 minutes (34.2 keyframes/sec on single NVIDIA GPU)
Integrity Checks	0 NaNs, 0 Infs, 100% L2 unit length verified

---

📄 License & Attribution

• Embeddings released under the Apache-2.0 License.
• Based on the Qwen3-VL foundation model by Alibaba Cloud & Qwen Team.