How google-deepmind/kinetics-i3d works
Project type and purpose
• This is an open-source research codebase providing pre-trained Inflated 3D ConvNet (I3D) models for video action recognition on the Kinetics dataset.
• It supplies model definitions, checkpoints, evaluation scripts, sample preprocessed data, and tests.
File structure
• Top level:
• i3d.py ‑ model architecture (Sonnet + TensorFlow)
• evaluate_sample.py ‑ example inference pipeline
• multi_evaluate.sh ‑ batch runner for different checkpoint combinations
• data/ ‑ checkpoints (rgb and flow streams, ImageNet-initialized and scratch), sample .npy arrays and label maps
• out/ ‑ text outputs from multi_evaluate.sh
• No separate “frontend” or “backend” – this is a library/tool for model inference & evaluation.
• Key external configuration: use of TensorFlow, DeepMind Sonnet, NumPy, MediaPipe/OpenCV for preprocessing.
README for additional insights
• Two-stream I3D architecture: separate RGB and optical-flow networks, combined by adding logits.
• Sample workflow:
a) build TensorFlow graph (calls i3d.InceptionI3d)
b) restore checkpoint from data/checkpoints
c) load preprocessed .npy video clips
d) run session to obtain logits and probabilities
e) output top-20 classes
• multi_evaluate.sh automates inference over four variants (RGB/flow × ImageNet-initialized/scratch) and writes results in out/.
Elements
• Blue for data stores
• Green for processing modules (preprocessor, model builder)
• Orange for orchestration scripts
• Gray for external dependencies