RecapNet

RecapNet: Action Proposal Generation in Untrimmed Video Sequences

RecapNet generates temporal action proposals from untrimmed video sequences by modeling short-term contextual dependencies and ranking candidate segments based on actionness probability.

Key Features:

• Residual Causal Convolution Module: Constructs short-term memory of past events using residual causal convolutions to preserve temporal context.
• Joint Probability Actionness Density Ranking: Computes and ranks candidate action segments using joint probability estimates of actionness density.
• Single-Pass Processing: Processes full-length video sequences in a single forward pass to generate comprehensive action proposals.

Scientific Applications:

• Action Detection: Improves temporal localization performance in untrimmed video analysis.
• Video Analytics: Supports surveillance, sports analytics, and human–computer interaction through automated action proposal generation.

Methodology:

Integrates residual causal convolution–based temporal memory with probabilistic actionness density ranking to identify and prioritize relevant action segments. Performance validated on benchmark datasets including THUMOS14 and ActivityNet-1.3.