Caching Segment Context¶
Utilities for aggregating segment-level context that supplements frame-level data.
Aggregators run after the per-frame conversion stage and operate one
segment at a time, reading the segment’s npz files, computing the
context (past / future trajectories, aggregated detection trajectories,
…), and writing each frame back in place. Segments are independent —
disjoint files, no shared aggregator state — so the per-segment loop
fans out across a multiprocessing.Pool controlled by the same
--num_workers and --do_parallel_processing flags as the frame
stage. The default start method is "spawn" because the parent
process imports TensorFlow / OpenCV at module load and fork would
inherit that state into workers.
Note
Custom SegmentContextAggregator
subclasses must be picklable: a spawn worker re-imports the
class by qualname, so the class needs to live in an installed
package (not an ad-hoc test module) and avoid carrying
un-picklable state (open file handles, locks, GPU contexts). The
escape hatch is to call process(..., do_parallel=False) for a
sequential fall-back.
|
Abstract pass to add history/future context back into transformed frames. |
|
Derive past/future ego-relative trajectories from pose matrices. |
|
Aggregate future detections into the current frame's modality payload. |