The use of high bit-rate multimedia sensors in networked applications poses a number of scalability challenges. In this paper, we present how IRISNET, a software infrastructure for authoring wide-area sensor-enriched services, supports scalable data collection from such sensors by greatly reducing the bandwidth demands. The architecture makes a number of novel contributions. First, it enables the use of application-specific filtering of sensor feeds near their sources and provides interfaces that simplify the programming and manipulation of these widely distributed filters. Second, its sensor feed processing API, when used by multiple different services running on the same machine, automatically and transparently detects repeated computations among the services and eliminates as much of the redundancy as possible within the soft real-time constraints of the services. Third, IRISNET distinguishes between the trusted and untrusted services, and provides mechanisms to hide sensitive sensor data from the untrusted services. Using implementations of a number of real world sensor-enriched services on IRISNET, we present an evaluation of the benefits of our distributed filtering architecture. Our evaluation shows that our design can: 1) reduce the bandwidth demands of many applications to a few hundred bytes per second on each sensor, 2) support a large number of services on each sensor through the use of redundant computation elimination, and 3) address privacy/security concerns with little additional overhead.