Abstract: Millions of video cameras are deployed globally across major cities for learning-based video analytics (VA) applications, such as object detection. The video streams from the cameras are typically sent over the wide-area network to be (at least partially) processed by one or more servers in a cluster or cloud. In this paper, to reduce the network latency and minimize reliance on the cloud, we explore using edge nodes closer to the camera to fully serve the video processing needs of the VA application. Given the resource-constrained nature of edge nodes, an edge-only solution necessitates disaggregating the VA application pipeline across available edge nodes. In the context of such disaggregated VA pipelines, we present OVIDA, a VA orchestrator that leverages available edge resources to host the VA application. OVIDA horizontally scales and places the disaggregated modules of the VA pipeline across edge nodes to maintain the frame rate required for performance while minimizing resource needs. To adapt to varying resource availability and reduce queueing delays, OVIDA implements a placement heuristic and a central queue architecture. To account for dynamic and heterogeneous network conditions while load balancing, OVIDA implements a lightweight, sampling-based reinforcement learning load balancer. We implement OVIDA in Kubernetes and evaluate its performance for containerized VA applications. We find that OVIDA can increase frame processing rate by a factor of 1.93 and reduce frame drops by about 90\% compared to a GPU-aware placement policy.