Mobile Edge Computing (MEC) with offloading from local light models to edge heavy models offers a promising approach to meeting the high processing quality and low latency demands of video analytics while reducing overall system resource consumption. However, significant challenges remain in predicting offloading gains and ensuring effective multi-device collaboration under dynamic video content and network conditions, as existing methods struggle with accuracy prediction, timeliness, and resource allocation flexibility. In this paper, we present a Tile Confidence Gains-inspired Vickrey Auction-based Offloading System (TCGVA), designed to optimize latency-decayed offloading accuracy gains under network and computation constraints. To dynamically and efficiently predict offloading improvements, the system leverages historical gains thresholds feedback from edge batch processing, incorporating recent offloading intervals. A distributed Vickrey auction-based strategy is proposed to optimize resource allocation among multiple devices, enabling quick adaptation to content and network fluctuations. Our implemented system, tested on the VisDrone dataset with HSDPA network traces, achieves 93.28% of the ideal accuracy gain while maintaining prediction speeds close to the greedy method, with 62.36% less bandwidth occupancy and 16.49% fewer latency violations under fluctuating conditions.