[Paper Review] Playing for Data: Ground Truth from Computer Games

Stephan R. Richter, Vibhav Vineet+2 more|arXiv (Cornell University)|Aug 7, 2016

Advanced Neural Network Applications30 references151 citations

TL;DR

The paper demonstrates how to extract pixel-accurate semantic labels from off-the-shelf computer games by intercepting rendering communications, creating a large labeled dataset, and showing that game data improves real-world semantic segmentation models and reduces manual labeling effort.

ABSTRACT

Recent progress in computer vision has been driven by high-capacity models trained on large datasets. Unfortunately, creating large datasets with pixel-level labels has been extremely costly due to the amount of human effort required. In this paper, we present an approach to rapidly creating pixel-accurate semantic label maps for images extracted from modern computer games. Although the source code and the internal operation of commercial games are inaccessible, we show that associations between image patches can be reconstructed from the communication between the game and the graphics hardware. This enables rapid propagation of semantic labels within and across images synthesized by the game, with no access to the source code or the content. We validate the presented approach by producing dense pixel-level semantic annotations for 25 thousand images synthesized by a photorealistic open-world computer game. Experiments on semantic segmentation datasets show that using the acquired data to supplement real-world images significantly increases accuracy and that the acquired data enables reducing the amount of hand-labeled real-world data: models trained with game data and just 1/3 of the CamVid training set outperform models trained on the complete CamVid training set.

Motivation & Objective

• Motivate large-scale, pixel-accurate semantic labeling for outdoor scenes beyond manual datasets.
• Show that commercial games provide diverse, photorealistic data for training segmentation models.
• Develop a detouring-based pipeline to extract persistent object signatures without access to game source code.
• Automatically propagate labels across frames and instances to accelerate annotation.
• Demonstrate that game-derived data improves real-world segmentation performance and reduces required real-world labels.

Proposed method

• Intercept the game-to-GPU rendering pipeline via detouring to record rendering resource usage.
• Hash and persistently identify meshes, textures, and shaders to create stable object signatures across sessions.
• Render two passes per frame: a conventional color pass and a second pass encoding per-pixel mesh/texture/shader IDs.
• Decompose images into patches defined by shared MTS (mesh-texture-shader) to form semantic units.
• Use association-rule mining to automatically propagate labels across patches sharing resources.
• Interactive annotation interface with rapid label propagation that reduces manual labeling time.

Experimental results

Research questions

• RQ1 Can pixel-accurate semantic labels be produced for images from commercial games without access to source code?
• RQ2 Does synthetic game data improve semantic segmentation performance on real-world outdoor datasets?
• RQ3 How much real-world labeling effort can be reduced by leveraging game-derived annotations?
• RQ4 How effectively can labels propagate across time and across instances using persistent rendering-resource signatures?
• RQ5 What is the diversity and realism impact of game-generated data on training robustness?

Key findings

• Labels for 24,966 GTA5 frames were produced in 49 hours, labeling 98.3% of pixel area.
• Pre-annotated areas via propagation covered 98.3% of the dataset and 73% of cumulative pixel area through MTS and rules.
• Using game data with 1/3CamVid training data improves mean IoU on CamVid by 3.9 percentage points.
• A model trained with game data plus 1/3 CamVid outperforms a model trained on the full CamVid training set.
• On KITTI, game data yields a 2.6 percentage point improvement over real-data-only training.
• Across experiments, synthetic game data augmented training outperforms real data-only baselines while reducing manual labeling burden.