README.md

---
license: mit
---

# Tencent Fall Detection: test videos and reference results

Six 1 fps test videos from two rooms, with the complete output of the
fall-detection and daily-report system for each. **Only days 6 and 11
contain a fall**; the other days have none.

| Video | Duration | Report generation | Room | Report date | What it shows |
|---|---|---|---|---|---|
| `day_1_1fps.mp4` | 2 h 47 min | 11 min | 1 | 2026-08-06 | Resident A (the resident) and Resident B (household member) with visitors |
| `day_6_1fps.mp4` | 3 h 30 min | 11 min | 1 | 2026-08-07 | **a fall** by Resident A; two visitors |
| `day_11_1fps.mp4` | 8 h 12 min | 17 min | 1 | 2026-08-08 | **a fall** by Resident A; one brief visitor |
| `day_7_1fps.mp4` | 8 h 48 min | 25 min | 1 | 2026-08-09 | Resident A with visitors |
| `day_8_1fps.mp4` | 8 h 03 min | 20 min | 1 | 2026-08-10 | Resident A with visitors |
| `day_19_1fps.mp4` | 7 h 46 min | 24 min | 2 | 2026-08-11 | the face-recognition fallback: nobody enrolled for face recognition appears, so the person present longest becomes the resident, shown as "the resident" (not "Resident A") |

Report generation is the VLM segment summaries plus the language-model
summary and report, model loading included, on one NVIDIA RTX 5090. Face
recognition runs alongside on the CPU and does not lengthen it.

`results/<video name>/json_summary/daily_report.json` is the system's final
output (overview, timeline, narratives, routine comparison, verified falls).
The rest is supporting evidence: `verify/` (fall-alarm verdicts and clips),
`stgcn/` (detector overlay videos and the presence CSV) and the other
`json_summary/` files (per-segment observations and the language model's
input). Face-recognition data is not published. Each room-1 report compares
its routine with the earlier dates (day 1, the first, has none); day 19
(room 2) has no baseline.

The reports were regenerated on 2026-10-01 with the current report code
(GitHub `feature/identity-linker`, commit `d4d52a1`: fall times read on the
right clock, visit edges fitted to the detector, the get-up time after a
fall, company the detector does not show removed, and other fixes). Only the
language-model summary and report were re-run: the detector output, the
fall-alarm verdicts (re-run: the same answers), the clips and the VLM segment
summaries are the earlier run's, except that `verify/verdicts.json` now gives each confirmed fall's
time on the video's own clock (`fall_time_absolute`). The report-generation
times above are the earlier full runs'.

## Alternative models: Nemotron throughout

`results_nemotron/<video name>/` holds the same six days with Nemotron 3 Nano
Omni 30B-A3B Reasoning (NVFP4) in place of every Qwen model: fall verifier,
segment-summary VLM and the language model of the summary and report. Made
with GitHub branch `feature/vlm-variants` (commit `1ab233f`), the stages
`VERIFIER=nemotron MAP_VLM=nemotron SUMMARY_LLM=nemotron bash
bash/run_video_system.sh` runs, on the same detector output, alarm events and
face-recognition data as `results/`. Each folder has `json_summary/` (the
summary, the report and the VLM's per-segment answers), `verify/` (the alarm
events and Nemotron's verdicts) and `action_classes.json`; detector files and
clips are those of `results/`.

| Six days | Qwen (`results/`) | Nemotron (`results_nemotron/`) |
|---|---|---|
| Real falls confirmed (of 2) / false falls | 2 / 0 | 2 / 0 |
| Real visitors found (of 14) / extra people listed | 14 / 0 | 14 / 4 |
| Resident's time in the room, mean error per day | 3.1 % | 5.8 % |
| Resident's glasses described right (days) | 6 / 6 | 4 / 6 |
| Better daily summary in a blind comparison (days) | 3 / 6 | 3 / 6 |

Nemotron describes one person's clothes, glasses or hair differently from
one 10-minute segment to the next, so a visitor can be listed twice
(days 1, 6, 19) and part of the resident's day can appear as a visitor
(day 8). Its report calls the woman resident of day 19 "he". Qwen remains
the default.