FOI’s WiFireNet study reported handgun-class average precision of 72.5% at a 25% intersection-over-union threshold and 53.8% at 50% for its WiFi-regular configuration. Both scores measure the same object class and setup at different box-overlap thresholds, in controlled indoor experiments with moving groups. (FOI study)
What the two scores measure
Average precision (AP) summarizes detection performance for a class. Intersection over union (IoU) measures how much a predicted bounding box overlaps the reference box; a higher threshold requires greater overlap. So the 72.5 and 53.8 figures are WiFi-regular’s handgun AP at two IoU thresholds—not scores for two different detection tasks.
| Study configuration | Handgun AP at 25% IoU | Handgun AP at 50% IoU |
| WiFi-plus, large-room setup | 44.2% | 44.2% |
| WiFi-regular, large-room setup | 72.5% | 53.8% |
| WiFi-plus, small-room setup | 40.4% | 40.4% |
| WiFi-plus, large-room subset with no more than two people | 29.6% | 27.3% |
These are handgun-class AP scores for the listed configurations. The study also reports overall mean average precision across all classes, a separate metric from handgun AP.
The rooms and data behind the results
The larger test area covered 12 m² and included up to four people moving through it. The separate 8 m² setup included up to two. At most one person in a scene carried a concealed, deactivated metal handgun. The other object classes were a plastic handgun, a metal water bottle and a mobile phone.
The researchers collected around 100,000 channel-state-information (CSI) samples for each of the WiFi-plus and WiFi-regular configurations, and 40,000 for the small-room setup. CSI describes changes in a Wi-Fi radio channel as signals travel through a space. Data came from a modified open-source IEEE 802.11a implementation using GNU Radio, operating at 5.68 GHz with 20 MHz bandwidth.
The split was made at the individual CSI-sample level: half the samples went to training and half to testing, with 20% of the training portion reserved for validation. The study’s scores therefore come from that sample-level evaluation and those bounded room setups.
Cameras in training, Wi-Fi at inference
The researchers recorded RGB video at 30 frames per second to annotate the data and train a vision-based teacher model. CSI bursts were sampled at 80 Hz. The paper says WiFireNet uses CSI alone when making inferences; camera footage played a role in preparing and training the research system, not as an inference input.
