Improving Object Detection by Modifying Synthetic Data with Explainable AI

Nitish Mital, Simon Malzard, Richard Walters, Celso M. de Melo, Raghuveer Rao, Victoria NocklesarXiv 2024 · DOI · arXiv

Introduction

In specialized domains such as defense, healthcare, and autonomous driving, the scarcity of high-quality real-world data presents a significant barrier to training robust object detection models. While synthetic data generated by 3D engines like Unity offers a scalable solution, it introduces a "Goldilocks" problem: determining the optimal level of realism. Simply increasing visual fidelity does not always improve performance; models often learn spurious features from synthetic data, leading to confusion between similar classes. This research proposes a data-driven framework that leverages Explainable AI (XAI) to replace manual intuition with a systematic roadmap for modifying synthetic training sets.

The XAI-Guided Framework

The core of this approach is a six-step iterative process centered on SHAP (SHapley Additive exPlanations) saliency maps. Unlike traditional workflows where XAI is used post-hoc for model validation, this framework utilizes XAI as a diagnostic tool for the dataset itself.

The six-step iterative framework using SHAP saliency maps to guide synthetic data modifications.
  1. Initial Training: An object detector (YOLOv8) is trained on a combination of real-world images and baseline synthetic data.
  2. Evaluation: Performance is assessed using a confusion matrix to identify specific "hard samples" and persistent class confusions.
  3. XAI Analysis: SHAP maps are generated for both correct and incorrect classifications to visualize the pixels the model prioritizes.
  4. Feature Identification: Researchers compare these maps to distinguish between Unique Features (which aid correct classification) and Common Features (which cause inter-class confusion).
  5. Targeted Modification: The 3D mesh models in Unity are adjusted using two strategies: Reinforcing Modification (enhancing unique features) or Disruptive Modification (reducing the prominence of confusing features).
  6. Retraining: A refined synthetic dataset is generated, and the model is retrained to evaluate improvements.

Experimental Setup and Datasets

The study utilized the ATR DSIAC infrared dataset, containing mid-wave infrared (MWIR) videos of eight vehicle classes. To rigorously test the model's generalization capabilities, the authors employed an "unseen orientation" split. The model was trained on real vehicles facing specific directions and then tested on orientations entirely absent from the real-world training set. Synthetic data was generated in Unity3D to fill these orientation gaps, with material properties like reflectance and thermal emission controlled via the Probuilder tool.

Data split strategy showing real-world training orientations versus unseen test orientations, and how synthetic data fills the gaps.

Key Results and the "Realism Paradox"

The researchers evaluated the framework using the mean Average Precision at 50% IoU (mAP50) metric. The baseline model, trained on real data only, achieved 90.0% mAP50. Adding basic synthetic data (v0) improved this to 94.6%. However, the XAI-guided modifications pushed the performance to 96.1% mAP50.

A critical insight from this study is the "Realism Paradox": making data less realistic can sometimes be more beneficial than increasing fidelity. For instance, by darkening a "hot" engine panel on a synthetic ZSU23 vehicle—making it less realistic compared to a true infrared signature—the researchers successfully disrupted a feature that previously caused the model to confuse it with a BTR70. This suggests that XAI can effectively "steer" a model's attention toward robust, essential features rather than noise or shortcuts.

Conclusion

This framework demonstrates that XAI can serve as a powerful debugger for synthetic datasets, providing a 6.1% performance boost over real-world baselines. While the current process requires a human-in-the-loop to interpret SHAP maps and modify 3D meshes, future work aims to automate this pipeline. By mapping 2D SHAP hotspots back to 3D mesh coordinates via ray-tracing, it may be possible to programmatically adjust textures and geometry, further streamlining the creation of high-performance synthetic training data.

Summary updated

BibTeX
@misc{mital2024improving,
  author = {Nitish Mital and Simon Malzard and Richard Walters and Celso M. de Melo and Raghuveer Rao and Victoria Nockles},
  title = {{Improving Object Detection by Modifying Synthetic Data with Explainable AI}},
  year = {2024},
  eprint = {2412.01477},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.01477},
}
Figure