PlaceTally builds a place graph before any question, so a returning camera stops double-counting the same room
Synopsis
When a camera returns to rooms it has already seen, a vision-language model reading the whole video answers with a few small numbers, while adding up parts counts the same object again in every part that shows it. PlaceTally builds a place graph for each video before any question is asked, grouping fragments by whether the camera has been here before, then lets a frozen vision-language model count each place once from at most 128 of its frames. On RevisitCount, a benchmark of real room re-recordings, counting by place reaches 62.5% mean relative accuracy on test rooms against 51.2% for whole-video reading with the same model, and 61.2% against 44.3% on held-out rooms evaluated once.
Interpretation
The unit of counting moves from the whole video or loose fragments to the place: when the camera returns to a room, the same objects are counted once for that place, so the method need not know in advance which category will be asked. Whole-video reading collapses onto a few small integers, with 87% of its test answers being 1, 2 or 3, while summing over parts recounts an object in every part that shows it; Cambrian-S's surprise memory predicts a median of 8.6 times the true count on these videos. On the ARKit-Revisit test split of RevisitCount, whole-video reading with the same frozen Qwen3.5-9B reader scores 51.2% and PlaceTally scores 62.5%; on the held-out split evaluated once after the method was fixed, the figures are 44.3% and 61.2%.
The place graph is built before any question, so one memory serves any category without a category vocabulary and without per-question object tracking. Existing tracking methods need the target first: CountVid needs the category before the video is processed, StaMina registers targets when the stream starts, and VideoAgent and DirectMe rely on a detector vocabulary. Each fragment makes one decision on its strongest earlier match, continuing the current place, joining the place of that match, or opening a new place; place F1 is 96.9% on ARKit test streams and 75.3% on 3RScan test streams.
At most 128 frames are kept per place, and that single number sets the reader's input, the memory footprint and the answer time, with the graph built 15 times faster than real time. Whole-video reading uses 2.90 MB per minute, 69.1 thousand input tokens and 15.9 GPU-seconds per query; PlaceTally at 128 frames uses 0.52 MB, 19.0 thousand tokens and 2.5 GPU-seconds. At 128 frames PlaceTally scores 50.1% against 51.2% for whole-video reading; at 512 frames it scores 62.6%, stores about half as much and answers 2.2 times faster.
Perspective
The demonstrated setting is streams joined from real ARKitScenes room re-recordings, with four to six rooms per stream and about a third of visits being returns, over 14 categories of furniture, fixtures and appliances; 3RScan-Revisit tests the same assumption with rescans months apart, and MultiScan-Revisit is used only to evaluate places. The authors propose fitting a continuation model to each camera and training return evidence on look-alike rooms as next steps.
Places stand in for object identity only while objects stay in their rooms between visits, an assumption the authors tie to their 14 categories. The streams join real recordings at a hard cut, and the other return rates come from removing visits from tested rooms rather than from new recordings of continuous walks. A change of camera requires refitting the continuation model, which is why 3RScan with its own fragments falls below the true-visit result.
