wenmin-wu/cv-cross-view-count-consistency-filter

Overview

Multi-camera event detection (NFL impacts, surveillance, sports tracking) provides a free post-processing filter: when two cameras observe the same play from different angles at the same moment, every real event should produce detections in both views. False positives are usually idiosyncratic to a single camera (compression artifact, motion blur in one view only), so frames where the two views disagree on the box count are almost certainly contaminated. This is a one-pass groupby-and-drop that costs nothing but throws away a meaningful chunk of FPs without touching the detector.

Quick Start

import pandas as pd

drop_idx = []
for (gk, pid), _ in test_df.groupby(['gameKey', 'playID']):
    play = test_df.query('gameKey == @gk and playID == @pid')
    for idx, row in play.iterrows():
        f = row['frame']
        n_side = play.query('view == "Sideline" and frame == @f').shape[0]
        n_end  = play.query('view == "Endzone"  and frame == @f').shape[0]
        if n_side != n_end:
            drop_idx.append(idx)

test_df = test_df.drop(index=drop_idx).reset_index(drop=True)

Workflow

1. Group detections by (gameKey, playID) — the unit of synchronized capture
2. For each frame within the play, count detections in each view
3. If the per-view counts differ, drop all detections in that frame (not just the extras)
4. Apply this after the per-view detector inference but before score-based filtering
5. Optionally allow a tolerance of ±1 if your detector is noisy

Key Decisions

• Drop the whole frame, not the surplus: you can't tell which of N detections is the spurious one — if the count is wrong the frame is suspect end-to-end.
• Per-frame, not per-play: a real impact at frame 100 shouldn't be dropped because frame 200 is asymmetric.
• Strict equality, no tolerance: synchronized cameras at 60fps are well-aligned; a tolerance often hides real disagreements.
• vs. cross-view IoU matching: counting is O(1) per frame; geometric matching needs camera calibration that's usually not provided.
• Only useful when both views are detector inputs: if you only have one camera, this trick doesn't apply.

References

• Both zones 2Class Object Detection strict filter