Adaptive QP selection appears to be incorrect when multiple slices / slice segments are used.

The statistics are reset at the start of compressSlice, and the delta calculated previously is added to the QpBase (a rate-controlled version of the slice's QP, used only for Adaptive Qp Selection) if the slice is not an I_SLICE.

The statistics are used to update the delta at the end of encodeSlice.

However, all slices are compressed. Once all slices are compressed, they are then encoded. So the statistics will only ever represent values gathered from the last slice of the previously coded picture.

Is the intension to gather statistics on a picture and use that for the next picture, or to gather statistics on a slice (including all slice segments) and use that for the next slice?

In addition, the 52 element array TComTrQuant::m_qpDelta will underflow if negative QPs are used as an index (valid for bit depths > 8).