Weather changes outside objects can disrupt cooperative detection
Understanding and Repairing Cooperative Perception under Adverse Weather · arXiv:2610.35110
In plain words
Cooperative 3D detection combines measurements from vehicles and roadside sensors to find objects. This paper introduces AWCR, a method that filters suspected weather returns in simulations and adjusts or selects detection candidates on real inputs.
In controlled tests, changing observations outside objects caused some previously correct detections to fail even when the points inside the objects stayed the same.
AWCR improved car detection on the paper’s natural adverse-weather development data, compared with the detector’s original predictions.
The research
When vehicles and roadside sensors combine measurements, can a change in the weather observations outside an object undo a correct cooperative detection even if the detector’s retained points inside the object stay unchanged?
The paper tests whether exterior changes can remove help that a receiver could not provide on its own.
The paper introduces AWCR, a repair method that filters simulated weather returns or adjusts and ranks real detection candidates using current measurements.
Researchers kept the detector fixed and preserved its measured points inside objects while changing only exterior observations. This created errors on real clear inputs and, in simulation, directional removal caused more errors than equal-count random removal.
Restoring original features around objects recovered most failures, linking changed object representations to detection errors. AWCR then improved car detection on natural adverse-weather development data and vehicle detection on a simulated-fog subset.
The controlled real-input tests changed clear-weather observations, and the natural-weather results come from development data. Real-input repair components were trained with annotated observations and constructed exterior-loss variants, without naturally paired clear and adverse-weather recordings.
What the evidence shows
- When exterior changes matter
The paper studies Weather-Induced Context Interference: weather-related changes outside an object can alter how a detector represents preserved object measurements and remove cooperative detection help. The authors introduce Adverse-Weather Cooperative Repair (AWCR), a method for repairing cooperative 3D detections in simulated and real inputs.
Their central question is whether changing exterior observations alone can make a previously correct detection fail.
- Isolating exterior changes
Researchers kept detector weights fixed and preserved every detector-retained point inside annotated objects while changing only exterior observations. On real clear-weather inputs, directional removal made 31 of 150 previously correct target observations incorrect; 21 of those losses were cases where the receiver could not provide the lost help alone.
In the main simulated V2X-ViT setting, directional removal caused eight new errors among 216 originally correct observations, compared with two to four under equal-count random controls across three fixed seeds.
- Tracing the failure pathway
To trace how exterior changes could affect preserved points, the researchers replaced selected pre-fusion features with their versions from the original input and continued detection. Restoring features around objects recovered seven of eight simulated failures, compared with four when restoring the exterior complement.
On real controlled inputs, object-neighborhood restoration recovered 30 of 31 failures, while each of three equal-area exterior masks recovered none. The contrast supports an object-region representation pathway for the observed failures.
- Repairing detections
AWCR addresses the identified problem with different procedures for the two input settings. In simulation, Point Classifier filters added returns identified as weather.
On real inputs, Local Correction supplies bounded candidate alternatives, and Quality Selection scores candidates using current measurements from both sources. On natural adverse-weather development data, AWCR improved V2X-ViT’s car detections over its original predictions; it also improved vehicle detections on a fixed simulated-fog subset.
- Bounding the correction result
The Selection Only comparison helps bound the added value of corrected candidates. It rescores original candidates with the same observation-guided regressor but does not use Local Correction to supply alternatives.
On V2X-ViT’s natural adverse-weather CoInfra development data, AWCR scored one car [email protected] point higher. That comparison supports a benefit in this tested setting, rather than establishing the same contribution across other detectors or settings.
Deep
- Original paper authors
- Mengxu Ma
- Submitted to arXiv
- License
- arXiv non-exclusive