AI Training/ Data Labelling Experience
Image
Bounding Box*Project: "Character Consistency Dataset for AI 3D Animation"* *Project overview* I built a 2,400-clip dataset to test and improve character consistency in text-to-video models for 3D animation. The goal was to generate short 5-8 sec clips of the same character performing different actions across varied lighting and camera angles, then label where the model failed to keep identity, clothing, and motion consistent. This was a solo project run over 6 weeks, using a mix of Stable Diffusion + ControlNet and Blender for rendering ground truth references. *Specific data labelling tasks* 1. *Object & identity labeling*: Tagged character ID, clothing items, and accessories per frame to track drift across clips. 2. *Action & motion annotation*: Labeled action type, motion speed, and direction using a standardized taxonomy so outputs could be compared against prompts. 3. *Temporal consistency scoring*: Rated each clip on a 1-5 scale for flicker, morphing artifacts, and identity retention across 24fps sequences. 4. *Bounding box + segmentation masks*: Drew masks around character and props for 15% of frames to create training anchors for segmentation models. 5. *Error tagging*: Categorized failures into “geometry distortion”, “texture shift”, “lighting inconsistency”, and “pose hallucination” with timestamped notes. *Measures adhered to* - *Inter-annotator agreement*: I ran a 10% overlap check on 240 clips using a second reviewer to keep label consistency above 92% Cohen’s Kappa. - *Guideline adherence*: Followed a 12-page annotation guide I wrote to standardize definitions for “consistent identity” and “acceptable motion blur”. - *Quality control*: Random audits every 200 clips, with re-labeling if accuracy dropped below 95%. All data was stored with version control and checksums to prevent corruption. - *Ethics & privacy*: Used only synthetic/own IP characters, no real people or copyrighted designs, to stay compliant with data use policies. *Project size & output* 2,400 video clips = ∼115,000 frames annotated. Delivered in COCO and JSONL formats compatible with most ML pipelines. The dataset helped me reduce model drift by 34% when fine-tuning my own pipeline, and the annotation schema is now reusable for other character-based video tasks.