A method for recovering information about light that conventional cameras cannot record placed UC Riverside researchers among the 10 Best Paper Award nominees at the European Conference on Computer Vision (ECCV), one of the field’s premier venues.
The paper was part of a broader showing by researchers from UCR’s Marlan and Rosemary Bourns College of Engineering at ECCV 2026, held Sept. 8-12 in Malmö, Sweden. Three studies involving BCOE faculty and students explored different limits on what machines can perceive, including information hidden from conventional image sensors and objects outside an autonomous vehicle’s field of view.
ECCV brings together academic and industry researchers working in computer vision and machine learning. ECCV accepted 2834 papers out of 10,473 submissions this year. 163 papers (1.6%) were accepted as oral presentations (28 long orals, 135 spotlights), and 10 papers were selected as award candidates. BCOE researchers presented two long oral papers and one spotlight paper.
Computer vision focuses on how machines perceive and interpret visual information and is one of the foundations of artificial intelligence (AI). The BCOE work presented at this year’s conference approaches that problem at several levels, beginning with the light that reaches a camera.
“Provable and Robust Wavefront Sensing via Self-Reference Interferometry”, by Nebiyou Yismaw, Vishwanath Saragadam, Aswin C. Sankaranarayanan (CMU), and M. Salman Asif, was selected as one of 10 Best Paper Award nominees and as one of 28 long Oral papers.
The research addresses a limitation built into conventional cameras. Image sensors measure the intensity of incoming light, or how much light reaches each point on a sensor, but they do not directly capture its phase. Phase describes how a light wave travels through space and can be essential for producing high-quality images in systems used in astronomy, microscopy, and other forms of advanced imaging.
Existing techniques can recover phase by comparing incoming light with a separate reference beam. That approach works, but maintaining a stable reference beam can be difficult outside carefully controlled conditions.
Yismaw and his collaborators instead use the incoming light as its own reference. Their method compares the light wave with spatially shifted copies of itself, measures the differences among those copies, and reconstructs the full phase profile mathematically.
The researchers also developed a theoretical framework showing which measurement patterns make that reconstruction reliable. In simulations, they recovered complete phase profiles from as few as eight shifted measurements. A hardware prototype demonstrated optical phase recovery, automatic refocusing, and imaging through scattering material.
Details about the work can be found at https://csiplab.github.io/coprime-psi/.
A second oral presentation examined a different problem in imaging.
“Broadband Wide Field of View Imaging with Computational Mirrors”, by Vishwanath Saragadam, Niki Nezakati, Amit Roy-Chowdhury, and Vivek Boominathan (Rice University), combines simple reflective optics with computational image reconstruction.
Broadband cameras can record visible light as well as wavelengths that people cannot see, including near-infrared and short-wave infrared light. Those wavelengths can reveal information about materials and objects that does not appear in a normal photograph.
Focusing all of that light with conventional glass optics is difficult because different wavelengths bend differently as they pass through glass. Optical systems compensate by using multiple lens elements, which can add size, weight, and complexity.
Mirrors do not have the same wavelength-dependent focusing problem, but simple mirror systems introduce severe distortions away from the center of an image. Different areas of the scene also come into focus at different distances from the mirror.
The researchers developed a way to use those imperfect images rather than eliminate the problem entirely through additional optics.
Their system captures a small set of images at different focus positions. One exposure may be sharp near the center while another captures sharper detail near the edge. A computational model called SeidelConv characterizes how the mirror distorts light across the image and helps combine the measurements into one image that is sharp across the field of view.
Tests showed that three images were generally enough for the systems studied.
The researchers built 50-millimeter and 100-millimeter prototypes capable of imaging wavelengths between 400 and 1,700 nanometers without refocusing for each part of the spectrum. Their experiments demonstrated why that broader range can be useful.
A real plant and an artificial plant that looked similar in visible light appeared different in near-infrared images. Short-wave infrared imaging revealed a bruise on an apple that was difficult to see in visible light. The system also captured security information on a $20 bill that changed across spectral bands.
Rather than relying on increasingly complicated lens assemblies, the approach shifts some of the work into computation and could support more compact broadband imaging systems.
Details about the work can be found at https://codelab-ucr.github.io/computational-mirrors/.
The third BCOE study moves the problem from individual cameras to connected vehicles.
“CooperScene: Multi-Modal Cooperative Autonomy Benchmark with C-V2X Communication Characterization”, by Bo Wu, Ruoshen Mo, Justin Yue, Yanyu Zhang, Janice Nguyen, Guoyuan Wu, Roy-Chowdhury, Matthew J. Barth, and Hang Qiu, examines how autonomous vehicles can share sensor information when their own view is blocked.
A vehicle’s cameras and other sensors can only detect what falls within their field of view. Another car, a large truck, or a building can hide a pedestrian or approaching traffic.
Cooperative perception is designed to extend that view by allowing connected vehicles and roadside equipment to share what their sensors detect. In practice, however, those systems depend on wireless communication that can be delayed, constrained by bandwidth, or affected by changing conditions.
CooperScene was built to study that problem under real-world communication conditions.
The dataset includes scenes recorded at intersections, highway ramps, local streets, and parking areas. Three connected and autonomous vehicles and one roadside unit collected information using multiple sensing technologies and commercial cellular vehicle-to-everything (C-V2X) communication radios.
Researchers annotated 344,000 objects across 59,000 frames and synchronized information among the vehicles, sensors, and roadside infrastructure. The dataset also captures characteristics of the wireless connection, allowing researchers to evaluate cooperative systems without assuming that communication is instantaneous or perfect.
That distinction is important for systems intended to operate outside a laboratory, where a vehicle may need to make a decision even when information from another vehicle arrives late or incompletely.
Details about the dataset and benchmark results can be found at https://cisl.ucr.edu/CooperScene/.
Together, the studies show how limitations in machine perception can arise well before an artificial intelligence system makes a decision. Sometimes information never reaches a conventional sensor. In other cases, optical hardware distorts what the sensor records. On the road, useful information may exist but sit beyond a vehicle’s line of sight.
BCOE researchers are working on each of those points in the process, using advances in imaging, computation, and connected autonomy to give machines access to information they otherwise would not have.
“Having three oral presentations accepted to ECCV reflects the growing strength of UC Riverside's AI and computer vision research community,” said Vassilis Tsotras, a distinguished computer science professor and co-director of the RAISE@UCR Institute, in an earlier UCR story about the conference. “Our faculty and students are tackling challenging scientific problems while developing technologies that have the potential to improve medicine, transportation, and many other fields.”