🐯 Tiger200K: Manually Curated High Visual
Quality Video Datasets from UGC Platform
Xianpan Zhou
zhouxianpan@qq.com
Abstract
The recent surge in open-source text-to-video generation models has significantly energized the research community, yet their dependence on proprietary training datasets remains a key constraint. While existing open datasets like Koala-36M employ algorithmic filtering of web-scraped videos from early platforms, they still lack the quality required for fine-tuning advanced video generation models. We present Tiger200K, a manually curated high visual quality video dataset sourced from User-Generated Content (UGC) platforms. By prioritizing visual fidelity and aesthetic quality, Tiger200K underscores the critical role of human expertise in data curation, and providing high-quality, temporally consistent video-text pairs for fine-tuning and optimizing video generation architectures through a simple but effective pipeline including shot boundary detection, OCR, border detecting, motion filter and fine bilingual caption. The dataset will undergo ongoing expansion and be released as an open-source initiative to advance research and applications in video generative models.
Visualization
The video features a person sitting in what appears to be a studio or interview setting. The individual is wearing a black shirt and glasses, and they seem to be engaged in a conversation or interview. The background includes bookshelves filled with books, suggesting an intellectual or academic environment. The lighting is warm and focused on the person, highlighting their facial expressions and gestures as they speak. The overall atmosphere is calm and professional.
The video captures a dramatic scene of a volcanic eruption. Bright, fiery lava is seen flowing and bubbling intensely from the volcano's crater, illuminating the surrounding area with its intense orange and red glow. Thick plumes of smoke rise into the air, adding to the sense of power and intensity. The lava appears to be moving dynamically, creating a mesmerizing display of nature's raw energy. The overall atmosphere is one of awe and danger, highlighting the unpredictable and powerful forces at play during such an event.
The video shows two cats, an adult and a kitten, sitting on a small woven mat. The adult cat is orange and white, while the kitten is mostly white with some orange patches. They appear to be in a relaxed setting, possibly outdoors, as there is a metal container nearby and a blurred background that suggests an open area. The adult cat seems to be grooming itself, while the kitten looks up at it, possibly observing or interacting with the adult. The scene conveys a sense of calm and companionship between the two cats.
The video shows a red seaplane flying low over a large glacier in a serene, icy landscape. The glacier is surrounded by calm waters and rugged mountains in the background. The plane appears to be gliding smoothly, providing a stunning aerial view of the natural beauty of the area. The scene captures the contrast between the vibrant red of the plane and the cool blues and whites of the ice and water.
The video features a person dressed in a pink and white outfit, performing a graceful dance in a garden setting. The dancer is adorned with floral accessories and has long, flowing hair. The background includes blooming pink flowers and lush greenery, creating a serene and picturesque environment. The dancer's movements are fluid and expressive, adding to the overall beauty of the scene.
The video captures a serene forest scene with a river flowing through it. The camera pans slowly, revealing tall trees with lush green foliage and some bare branches. The water reflects the surrounding trees, creating a mirror-like effect. Sunlight filters through the leaves, casting dappled shadows on the ground and water. The overall atmosphere is peaceful and natural, with the gentle flow of the river adding to the tranquil ambiance.
Pipeline
The pipeline of data construction. First, the selected video will segment by scene and further subdivided into cuts of fix length. Then, methods such as OCR, border detection, and optical flow estimation are used for quality filtering. Finally, manual review was carried out and bilingual fine-grained caption was performed using VLM.
Statistical results of video level
Videos with 4K and 1080P resolutions each account for approximately half. The vast majority of videos can only extract a small number of clips due to their short duration and strict filtering.
Statistical results of annotation level
The length of bilingual annotations is mostly controlled within a reasonable range. The retention area of safe zone for the vast majority of clips is also above 0.85.