What is TID and LGD?
TID and LGD are nuScenes tracking diagnostics that measure how long a ground-truth trajectory waits for its first successful match and the longest continuous interval for which it is missed.
Quick Facts
| Full Name | Track Initialization Duration and Longest Gap Duration |
|---|---|
| Specification | Official Specification |
How It Works
Measure first acquisition and the longest missed run
For one ground-truth track, TID is the elapsed time from its first evaluated occurrence to its first matched occurrence. LGD is the duration of the longest consecutive run of misses between the first and last evaluated occurrences. A match therefore depends on the benchmark's class filter, localization distance, assignment, ignore rules, and confidence threshold.
The nuScenes tracking protocol evaluates 2 Hz keyframes and reports TID and LGD in seconds. On another dataset, convert frame counts with real timestamps or the declared sample period; multiplying by 0.5 is specific to this protocol.
Pin the paper, Devkit, and operating point
The nuScenes paper states that a completely untracked object receives its full track duration for both metrics. The current reference Devkit instead excludes tracks with no match and averages over matched tracks. That difference can materially improve a weak submission's reported mean.
The standard summary computes these diagnostics at each confidence threshold, selects the per-class threshold with maximum clipped MOTA, and then averages class results. Report the Devkit commit, threshold-selection rule, matched-track coverage, and per-class distribution with the scalar.
Interpret delay and outage separately from identity
TID diagnoses initialization, confirmation-window, and early-detection failures. LGD diagnoses occlusion recovery, temporary detector failure, and track life-cycle termination. Neither identifies which predicted ID was used across successful matches, so an identity-swapping tracker can still have low values.
Use track completeness and fragmentation for coverage shape, IDF1 or HOTA association components for identity, and MOTA or AMOTA for false and missed detections. For safety analysis, also stratify delays by range, visibility, class, speed, and downstream time-to-collision rather than treating a dataset mean as a guarantee.
Key Characteristics
- Measures first-acquisition delay and longest continuous outage
- Uses seconds rather than a normalized accuracy scale
- Depends on benchmark matching and confidence threshold
- Exposes temporal failures hidden by aggregate recall
- Has a documented paper-versus-Devkit unmatched-track difference
- Does not directly measure identity consistency or localization error
Common Use Cases
- Auditing delayed initialization in autonomous-driving trackers
- Measuring recovery after occlusion or temporary sensor failure
- Tuning track confirmation and termination policies
- Comparing temporal reliability at one fixed operating point
- Building range, visibility, class, and speed failure slices
Example
Loading code...Frequently Asked Questions
How are TID and LGD calculated?
TID measures the time from a ground-truth track's first evaluated occurrence to its first accepted match. LGD scans that track for its longest consecutive run of unmatched occurrences. Both require the benchmark's matching and timing rules.
Are lower TID and LGD values better?
Yes. Lower TID means earlier acquisition, and lower LGD means shorter continuous outages. Compare them only under the same sample rate, class filters, matching threshold, confidence operating point, and treatment of never-matched tracks.
How does the nuScenes Devkit handle a never-matched track?
At commit b40adc4, it excludes a ground-truth track with no accepted match from both means and returns undefined if no tracks match. The paper's prose instead assigns the full track duration, so reports must identify the implementation.
What is the difference between LGD and fragmentation?
Fragmentation counts interrupted-and-resumed episodes, while LGD measures the duration of the longest missed run. A tracker can have one very long gap or many short gaps, so the two diagnostics answer different questions.
Do TID and LGD measure identity switches?
No. They use whether a ground-truth occurrence received an accepted match, not whether the same predicted identity persisted. Pair them with IDF1, HOTA association metrics, or raw identity-switch counts.