Counting distinct objects in a video is hard when the camera returns to places it has seen. A vision-language model that reads the whole video answers with a few small numbers, and one that reads it in parts counts an object again in every part that shows it. Object trackers need the category before the video is read, or a fixed vocabulary. We observe that a memory for counting does not need to know what it is looking for, only where it has been: objects rarely leave a room between visits, so parts of a video grouped by place share almost no objects, and a place can be recognised before any question is asked. Building on this, we introduce PlaceTally, a question-blind test-time memory model that constructs a place graph for each video. Each video fragment joins the place of its best earlier match or opens a new place, because a strong match is rarely wrong while a true return often matches weakly. A frozen vision-language model counts each place once from at most 128 of its frames, so every read has a bounded input and the memory grows with the number of places; the graph is built 15 times faster than real time. On RevisitCount, our benchmark of real room re-recordings, it scores 62.5% mean relative accuracy on test rooms and 61.2% on held-out rooms, against 51.2% and 44.3% for the same model reading the whole video. It stays level as rooms return more or less often, while a fixed correction of summed visits moves from 38.5% to 61.4%.
Figures & tables
Method
Identity of
Before question
Named in advance
Depth, path
CountVid [ 2 ]
target objects
no
the target
no
StaMina [ 45 ]
objects, events
yes
targets
no
VideoAgent [ 24 ]
objects
yes
detector
no
Emb. VideoAgent [ 25 ]
3D objects
yes
detector
yes
DirectMe [ 73 ]
objects, places
yes
Objects365 [ 62 ]
est.
Hydra [ 35 ]
objects, rooms
yes
labels
yes
Table 1: How methods decide identity. Depth, path: depth or camera path used (est.: estimated). Cells come from the primary papers. None of them uses place identity to stop double counting.
Figure 2: A PlaceTally memory is constructed per video instance before the question and read after it, shown on one real ARKit-Revisit test stream. (a) The detector cuts the stream into fragments. (b) Continuation and return evidence between fragments, which Eq. 2 would sum. (c) The place graph GV : Eq. 3 takes one decision per fragment (node), the kept links are its edges, and each lane is a place. Read: the category enters only here; each place is gated and counted, and the counts are summed once per place (Eq. 1 ).
Benchmark
Videos
QA
Length
Returns
Blind
VSI-Bench [ 76 ]
288
> 5,000
n/r
n/r
yes
VSI-SUPER-Count [ 77 ]
n/r
200
10–120 min
n/r
n/r
ReVSI [ 83 ]
381
n/r
n/r
n/r
partial
OST-Bench [ 42 ]
1,400
10,000
n/r
n/r
chance
EC-Bench [ 69 ]
152
1,699
57 min
n/r
n/r
TOC-Bench [ 14 ]
1,951
2,323
29 s
n/r
filter
Table 2: Benchmarks we inspected and RevisitCount (below the line). Returns: how a room is seen again. Blind: blind baseline reported (partial: object-removed videos; filter: only to filter questions). n/r: not reported. Our videos: streams with questions (MultiScan: all streams).
ARKit-Revisit
3RScan-Revisit
Method
test
held-out
test
Blind baselines † [ 30 , 15 ]
Fitted on training streams
41.6 [37.7, 45.4]
39.2 [35.4, 43.3]
37.1 [34.1, 40.1]
Rate × PlaceTally’s places ∗
39.8 [35.9, 43.4]
36.8 [32.9, 40.9]
32.0 [28.4, 35.8]
Adding up parts [ 56 , 77 ]
Sum over visits (privileged)
26.9 [23.5, 30.5]
30.8 [27.4, 34.3]
17.9 [14.6, 21.3]
Table 3: PlaceTally leads on real ARKit returns and falls behind whole-video reading under 3RScan’s camera change. Strict MRA (%) with 95% bootstrap intervals over streams. Privileged rows use true visits; bold: best row without them.
ARKit test (598 places)
3RScan test (824)
Variant
F1
Places
Exact
F1
Places
Exact
How the stream is cut (same linking)
Fixed 8 s windows
32.0
4393
0
59.5
1699
8
TransNet V2 [ 65 ]
80.7
510
27
48.8
423
6
SigLIP [ 68 ] similarity drop
85.7
764
32
70.9
2261
2
Detector, no continuation
96.3
764
51
67.8
1737
3
Table 4: Finding places, post hoc; F1 in %. Places: predicted (true in the header). Exact: streams with the right number of places, of all 109 and 150. Linking rows share log-odds weights, so Eq. 3 there differs from PlaceTally, which has its own thresholds. With MegaLoc, only the return evidence changes, its threshold fitted on training streams.
Method
Build speed
Memory MB/min
Input k tokens
Query GPU-s
MRA (%)
PlaceTally, K=16
15
0.52
19.0
2.5
50.1
PlaceTally, K=64
15
1.53
68.3
7.1
62.6
PlaceTally, K=128
15
2.32
105.8
11.1
62.5
Whole-video reading
–
2.90
69.1
15.9
51.2
VideoAgent [ 24 ]
0.88
15.5 ∗
n/c
15.3
4.9
CountVid [ 2 ]
–
7.6 ∗
–
11,084
0.0 †
Table 5: Cost on ARKit test (11-minute videos), post hoc, on one A100-40GB GPU. K : frames kept per place ( K=16 : refitted gate); PlaceTally and whole-video reading use Qwen3.5-9B. Build: × real time, once per video. Memory: kept after construction (PlaceTally: 3.0 MB/min while building). Input: all 460 questions; query: the same 50, nothing cached. –: none. n/c: counted differently. n/m: not measured. ∗ JPEG estimate (as run: 87, 68, 24 MB/min). ‡ GPU state while answering. † 40 questions.