zangxh/VORD

We propose VORD, the first evaluation dataset to simultaneously include video pairs (before and after removal) and graffiti masks. It consists of three types of video pairs including 🤖model-generated, 📸camera-captured, and 🛠️tool-rendered videos in equal proportions, covering 8 removal object types, 3 real-world scenarios, and 3 object association effects. Furthermore, compared to traditional segmentation masks, we introduce graffiti masks that better align with authentic user input. Specifically, we use manually drawn masks for the first frame to simulate real user interactions, while using segmentation masks for the subsequent frames. VORD contains a total of 150 test cases, where each case includes the original video (rgb_full.mp4), the ground-truth video after removal (rgb_removed.mp4), and the graffiti mask video (mask_drawn.mp4).