This paper describes an object-order real-time volume rendering architecture using an adaptive resampling scheme to perform resampling operations in a unified parallel-pipeline manner for both parallel and perspective projections. Unlike parallel projections, perspective projections require a variable resampling structure due to diverging perspective rays. In order to address this issue, we propose an adaptive pipelined convolution block for resampling operations using the level of resolution to keep the parallel-pipeline structure regular. We also propose to use multi-resolution datasets prepared for different levels of grid resolution to bound the convolution operations. The proposed convolution block is organized using a systolic array structure, which works well with a distributed skewed memory for conflict-free accesses of voxels. We present the results of some experiments with our software simulators of the proposed architecture and discuss about important technical issues.