TL;DR
This paper investigates a critical failure mode in object detection: "hallucination," where detectors falsely identify missing objects. They introduce DelftBikes, a novel dataset for visual part verification, and propose a new evaluation metric that prioritizes detecting present parts while penalizing false positives for missing ones, highlighting that standard object detection metrics overlook this issue.
What Problem Does This Paper Address?
Computer vision systems are increasingly used for tasks requiring precise verification, such as checking if all components are present in a manufactured product or if infrastructure is intact. A key challenge in these applications is "visual part verification," which involves determining the presence or absence of specific object parts. While standard object detection models are adept at finding existing objects, they can suffer from "hallucination" – falsely detecting objects that aren't actually there, sometimes even at their expected locations. This is particularly problematic for verification tasks, as a hallucinated "present" part could lead to incorrect assessments and potentially dangerous outcomes. Existing object detection benchmarks and metrics, like mean Average Precision (mAP), don't adequately capture or penalize this hallucination behavior, especially at lower confidence scores.
Methods
The authors first demonstrate the hallucination problem by evaluating popular object detection models (Faster RCNN, RetinaNet, YOLOv3) on their new dataset. They then propose a new evaluation framework for visual part verification. Instead of relying on standard precision and recall, they focus on a cost-weighted recall. This metric assigns a higher penalty to false positives (detecting a missing part as present) than to false negatives (failing to detect a present part). They also explore two evaluation setups: one with localization (requiring a certain Intersection over Union, IoU, threshold for detection) and one without localization (treating it as a full-image classification task).
Datasets
The paper introduces DelftBikes, a novel dataset specifically created for visual part verification. It comprises 10,000 images of bicycles, with 22 different parts densely annotated per image. Crucially, each part is labeled not only with its bounding box but also with one of four states: intact, damaged, absent, or occluded. This explicit annotation of missing or damaged parts is what distinguishes DelftBikes from existing object detection datasets. The dataset is split into 8,000 images for training and 2,000 for testing, with part state distributions kept similar across both sets.
Key Insights and Ideas
- Hallucination is a real and significant problem: Deep object detectors can indeed "hallucinate" missing objects, and they can do so with high confidence and accurate localization, making it difficult to detect this failure.
- Standard metrics miss the mark for verification: Metrics like mAP are not sensitive enough to penalize hallucinated detections of missing parts, especially when they occur at lower confidence levels.
- Visual verification requires different evaluation: The cost of a false positive (hallucinating a missing part) is often much higher than a false negative (missing a present part) in real-world verification scenarios. This necessitates a new evaluation approach that reflects these differing costs.
- Contextual and positional biases can be exploited: The inherent positional relationships between bicycle parts in the DelftBikes dataset might be exploited by detectors, leading them to predict parts that aren't there but are expected based on context.
Key Results
- Demonstrated Hallucination: All three tested detectors (Faster RCNN, RetinaNet, YOLOv3) exhibited hallucination, detecting missing parts with significant IoU scores.
- DelftBikes Dataset: Provides a valuable resource for studying visual part verification, with 22 densely annotated parts and explicit state labels (intact, absent, damaged, occluded).
- Evaluation Metric Performance: When evaluated with localization (IoU threshold of 0.5 for present parts, 0.1 for missing parts), Faster RCNN achieved the best visual verification score (Fvv) by detecting the fewest missing parts, outperforming RetinaNet and YOLOv3 despite similar object detection AP scores.
- Without Localization: Even when localization was ignored, a multi-class multi-label classification (MCML) approach performed comparably to Faster RCNN and better than RetinaNet and YOLOv3, suggesting that full-image context can be leveraged.
Limitations and Future Work
The authors acknowledge that the human annotations for non-existing parts are "partly guesswork." This makes the detectors' ability to predict these parts with high IoU scores even more surprising. Future work could involve investigating the role of contextual information in more detail to understand and mitigate hallucination. Additionally, exploring more sophisticated evaluation metrics that capture the nuanced costs of different types of errors in visual verification is a promising direction.
Summary updated