ramanv-video-foundation

A production-grade, highly-scalable dataset engineering platform designed for collecting, processing, validating, captioning, scoring, deduplicating, and publishing generative video datasets at scale.

The repository is built to process datasets scaling up to 1M+ videos, supporting keyframe extraction, multi-VLM captioning, OCR extraction, audio/speech transcription, CLIP/video embeddings, dense optical flow motion scoring, monocular depth mapping, semantic segmentation, deduplication, quality filters, and automated uploading to the Hugging Face Hub.

---

📂 Folder Layout

ramanv-video-foundation/
├── README.md                 # Dataset Card & Guide
├── LICENSE                   # Apache-2.0 License
├── pyproject.toml            # Project configurations and packages
├── requirements.txt          # Python dependency list
├── dataset_infos.json        # HF Dataset features schema
├── create_repo.sh            # Bash initialization script
├── create_repo.ps1           # PowerShell initialization script
│
├── configs/                  # Pipeline configurations (YAML)
│   └── default.yaml
│
├── videos/                   # Video folders grouped into 120 categories
│   ├── marketing_ads/
│   ├── product_ads/
│   ├── indian_weddings/
│   └── ... (117 other folders)
│
├── keyframes/                # Extracted keyframes (JPEG)
├── previews/                 # Low-resolution preview GIFs
├── thumbnails/               # Primary video cover images
├── captions/                 # Generated caption JSONs
├── metadata/                 # Schema-compliant JSON files (62 fields)
├── manifests/                # Dataset manifest files (JSON)
├── quality/                  # Quality reports
├── embeddings/               # Precomputed CLIP/Video embedding matrices (.npy)
├── optical_flow/             # Optical flow analysis files
├── depth/                    # Monocular depth maps
├── segmentation/             # Semantic segmentation masks
├── ocr/                      # OCR detections with timestamps
├── audio/                    # Extracted audio tracks (.wav)
├── transcripts/              # Transcripts with word timestamps
├── splits/                   # Train/Val/Test split records
├── logs/                     # Rotation log output files
└── src/                      # Core utility codebase

---

📊 Video Categories

The dataset organizes assets into 120 distinct category folders under videos/, including:

• Commercial & Ads: marketing_ads, product_ads, cosmetics, real_estate, fashion, food, restaurants, electronics, unboxing, reviews.
• Social Media: social_media, instagram_reels, youtube_shorts, vlogs, interviews, gaming, talk_shows.
• Culture & People: indian_weddings, festivals, temples, villages, city_life, street_markets, office, workspace, children, elderly, pets, dancing, cooking, diy.
• Visual Styles: cinematic, drone, timelapse, slow_motion, hyperlapse, stop_motion, animation, 3d, cgi, documentary, news, underwater, aerial, macro, first_person.
• Physical & Motion Elements: nature, water, fire, snow, space, sunset, night, sports, fitness, running, yoga.
• Industrial & Professional: science, technology, robots, medical, construction, agriculture, tutorials, e_learning, conferences, workshops.

---

🛠️ Getting Started

1. Clone & Install Dependencies:

   pip install -r requirements.txt
   

2. Initialize Folder Structure: On Linux:

   bash create_repo.sh
   

   On Windows (PowerShell):

   powershell -ExecutionPolicy Bypass -File .\create_repo.ps1
   

3. Add Raw Video Assets: Place your raw video files (.mp4, .mkv, .avi, etc.) inside their respective category subdirectories under videos/.

---

🚀 Sequential Execution Pipeline

To compile a fully processed and validated dataset, execute the processing scripts in the following sequence:

graph TD
    A[Raw Videos] --> B[Keyframe Extraction]
    A --> C[Audio Demuxing]
    B --> D[Visual Captioning]
    B --> E[Video OCR]
    B --> F[CLIP Embeddings]
    B --> G[Depth Estimation]
    B --> H[Semantic Segmentation]
    C --> I[Speech Transcription]
    A --> J[Video Embeddings]
    A --> K[Motion Analysis]
    D & E & F & G & H & I & J & K --> L[Quality Filtering]
    L --> M[Deduplication]
    M --> N[Create Splits]
    N --> O[Auditing & Validation]
    O --> P[HuggingFace Upload]

1. Keyframe Extraction

Extract representative keyframes, previews, and cover thumbnails:

python scripts/extract_keyframes.py --method adaptive --num-keyframes 5

2. Audio Extraction

Extract audio tracks (WAV formats are generated to feed transcription models):

python scripts/extract_audio.py --format wav

3. Speech Transcription

Run Whisper transcription with word-level timestamps and language identification:

python scripts/transcribe_audio.py --model base --gpu-id 0

4. Video Captioning

Generate short summaries and dense captions using Vision LLMs:

python scripts/generate_video_captions.py --model Qwen/Qwen2.5-VL-7B-Instruct --gpu-id 0

5. Video OCR

Extract onscreen text overlays and associate temporal timestamps:

python scripts/video_ocr.py --languages en

6. CLIP Embeddings

Extract sequence frame embeddings using OpenCLIP or SigLIP:

python scripts/compute_clip_embeddings.py --model-family openclip --model-name ViT-H-14

7. Video Embeddings

Extract unified spatial-temporal video representations:

python scripts/compute_video_embeddings.py --model microsoft/xclip-base-patch16

8. Motion Analysis

Compute Farneback dense optical flow to categorize camera and object movement:

python scripts/motion_analysis.py --sample-fps 5.0

9. Monocular Depth Estimation

Generate keyframe depth maps for visual structure mapping:

python scripts/depth_estimation.py --model LiheYoung/depth-anything-base-hf

10. Semantic Segmentation

Segment keyframes to isolate subjects, roads, sky, and backgrounds:

python scripts/segment_video.py --model nvidia/segformer-b0-finetuned-ade-512-512

11. Quality Filtering

Score visual fidelity, filter blurry/compressed clips, detect black screen errors, and screen watermarks:

python scripts/quality_filter.py --min-width 1280 --min-height 720

12. Deduplication

Cluster and remove exact or near-duplicate assets:

python scripts/deduplicate.py --phash-threshold 8 --cosine-threshold 0.96

13. Create Splits

Generate balanced training, validation, and testing partitions:

python scripts/create_splits.py --train-ratio 0.8 --val-ratio 0.1 --test-ratio 0.1

14. Dataset Validation

Audit metadata files and manifest schemas to ensure integrity:

python scripts/validate_dataset.py

15. Hugging Face Upload

Publish the compiled dataset folder to HuggingFace Hub:

python scripts/upload_hf.py --repo-id username/ramanv-video-foundation

---

🚀 Model Training Guide

1. Fine-tuning Video Diffusion Transformers (DiTs)

Modern video generative architectures (e.g., Sora, LTX-Video, CogVideoX) process video as a sequence of spatial-temporal patches.

• Dense Captioning: Always train using the long_caption metadata field. Large text encoders like T5-XXL parse spatial relationships, camera pan movements, and action sequences detailed in the caption.
• Temporal Aspect Bucketing: Group video training items by aspect ratio class and FPS bins to avoid spatial distortion and temporal jitter during batch training.
• Motion Filtering: Use the computed motion_score and camera_motion tags. To train high-action video models, filter out clips with motion_score < 0.2 (static scenes). For camera motion editing (e.g. pan left/right condition), filter using the camera_motion category labels.

2. Audio-Visual Video Generation

• Use the word-level transcripts in transcripts/ combined with the extracted audio track audio/ to train audio-visual synchronization networks, talking-head generators, or sound-aligned video generation.

3. Depth-conditioned ControlNet

• Train depth-conditioned spatial controllers by utilizing the computed depth maps under the depth/ folder as auxiliary controls, enabling structure-consistent video generation.

---

📄 License

This repository is licensed under the Apache License 2.0. See LICENSE for more details.