Research

Video in the open world never settles. New objects and scenes keep appearing, so a camera that was accurate the day it was deployed will not stay accurate. Labels are billed by the hour, footage cannot leave the node that captured it, and inference load swings from overload to idle across neighbouring nodes.

Our work closes a loop over three questions — what the edge should learn, how that knowledge spreads, and how inference actually runs — so that distributed edges keep adapting instead of quietly degrading.

01

What should the edge learn?

Value-Driven Data Acquisition

Video is overwhelmingly redundant: the same person is caught by several cameras, stays in frame for tens of seconds, and is often already handled correctly by the current model. Strip those three redundancies away and only about a tenth of the frames carry genuinely new knowledge — yet every label costs money. We learn a single acquisition policy, shared across edges, that scores each frame in its spatio-temporal context and keeps only what will actually move the model. Gradients leave the node; footage does not.

Edge 1Edge 2Edge 3shared policyscores each frame in context~30% kept & labeledmodel keeps adaptingGradients leave the node. Footage does not.
  • Reinforcement-learned frame selection under a labeling budget
  • Cross-edge redundancy suppression without pooling data
  • Rewards that balance accuracy, diversity and coverage

Under a 30% budget, ahead of every active-learning baseline we compared against.

AVA (AAAI’25)
02

How does knowledge spread?

Federated Continual Learning

Knowledge at the edge is spatio-temporally correlated — a person seen at one camera reappears later, and at the camera next door. An edge that has already adapted to a drifting scene is therefore worth listening to, except its images cannot be shared and naive model averaging just makes edges interfere. We freeze the extractor so images become compact prototypes, and let the server aggregate only the adaptive layers, weighted by how related two edges’ task histories are in space and time.

server: aggregate by task similarityin space and timeedge 1  ·  frozen extractoradaptive layeredge 2  ·  frozen extractoradaptive layeredge 3  ·  frozen extractoradaptive layer↑ parameterspersonalised model ↓Prototypes and parameters travel; images never leave the node.
  • Personalised aggregation by spatio-temporal task similarity
  • Prototype replay against catastrophic forgetting
  • No task identifiers required at deployment

Best accuracy among federated and lifelong baselines, with 62% less communication.

FedSTIL (TCSVT’25)
03

How does inference run?

Collaborative Inference Across Edges

Load is never balanced: one node queues up while its neighbour idles. Every request also carries three coupled decisions — how far to downsample it, which model should run it, and whether to keep it or forward it — and tuning any one alone leaves performance on the table. Each node runs its own agent on purely local observations and learns the joint action under a shared reward, so nodes cooperate without a central controller and without talking at inference time.

edge 1agentresolution · model · routerequestsedge 2agentresolution · model · routerequestsedge 3agentresolution · model · routerequestsEach agent sees only local state; all share one reward.
  • Joint control of resolution, model choice and request routing
  • Centralised training, decentralised execution
  • Attention over peer state as the cluster grows

33.6–86.4% better overall performance, 92.8% fewer dropped frames.

EdgeVision (TMM’24)