Mainstream adaptively merges the video stream processing of concurrent applications sharing fixed edge resources to maximize aggregate result quality. Mainstream’s approach enables partial-DNN compute sharing among applications using DNNs (deep neural networks) that are fine-tuned from the same base model, decreasing aggregate per-frame compute time. Moreover, since the choice depends on the mix of applications running on an edge node, Mainstream automatically determines at deployment time the right trade-off between specializing more of a DNN, which improves per-frame accuracy, and specializing less, which can increase sharing and thereby allow processing of more frames per second. Experiments with several datasets and event detection tasks on a mini-PC confirm that Mainstream provides huge reductions (often over 90%) in mean F1 score relative to the common approach of using full independent per-application DNNs (i.e., “no sharing”) or a static approach of retraining only the last DNN layer and sharing all others (“max sharing”).