What follows is a testing framework. It is written to be run as a structured pilot on your own footage, and it is deliberately agnostic about vendor. Most of these tests can be executed in a two-week evaluation with a small team.
1. Native format ingest
Test: Bring exports from every DVR and NVR brand your jurisdiction actually encounters – including the cheap unbranded units common in small commercial premises – and attempt direct ingest without conversion.
Why it matters: Conversion is the most common point of evidential loss in Indian casework. A platform that requires you to transcode before analysis pushes that loss into your standard workflow.
2. Metadata and timestamp handling
Test: Ingest a file with embedded timestamps and a known DVR clock offset. Check whether the platform preserves original metadata, exposes it to the analyst, and allows an offset correction to be applied and documented rather than silently overwritten.
Why it matters: Multi-camera timelines are built on this. Silent normalisation of timestamps is a defect, not a convenience.
3. Hashing and integrity verification
Test: Confirm that files are hashed on ingest, that hashes are recorded in an exportable log, and that the platform can re-verify an original at any later point.
4. Non-destructive processing with a reproducible log
Test: Apply a chain of six or seven processing operations to a working copy. Export the processing log. Hand it to a second examiner and ask them to reproduce the output independently.
Why it matters: If they cannot reproduce it, neither can your expert witness. This test alone eliminates a surprising number of otherwise capable products.
5. The generative boundary
Test: Take genuinely unresolvable footage – a face at ten pixels, a saturated number plate – and run every enhancement feature the platform offers. Examine whether any of them produce sharp, plausible detail. Our guide on what enhancement can legally and technically achieve covers this boundary in detail.
Why it matters: If the platform will generate a face from nothing, an analyst under pressure will eventually use it. Ask the vendor directly and in writing which of their operations are generative and which are strictly signal-recovery. A vendor that cannot answer this crisply does not understand the forensic market.
6. Multi-frame integration quality
Test: Select a stationary number plate visible across twenty or more frames at marginal legibility. Compare the platform’s multi-frame result against a single best frame.
Why it matters: This is the operation that produces real, defensible gains. Implementations vary enormously in quality.
7. Search and triage at volume
Test: Ingest at least one hundred hours across a dozen sources. Measure indexing throughput on the hardware you would actually buy – not the vendor’s demonstration server. Then run realistic attribute queries and measure both recall and false positive rate against ground truth you establish manually on a sample.
Why it matters: Throughput claims are usually quoted for ideal footage on unrepresentative hardware. Recall on your night-time footage is the number that determines whether the tool is useful.
8. Cross-camera association behaviour
Test: Follow a known subject across five cameras with varying angles and lighting. Record how many correct associations the platform proposes and how many false ones.
Why it matters: Over-eager association is worse than none, because it produces confident wrong timelines.
9. Deployment model and data sovereignty
Test: Establish precisely where processing occurs, where data resides, whether the system functions fully air-gapped, and what the platform transmits externally – including licensing checks and telemetry.
Why it matters: For most agency casework, evidence leaving the network is not a preference issue. Verify with a network capture during the pilot rather than accepting a statement.
10. Audit trail and access control
Test: As an administrator, reconstruct from logs which analyst opened which case, what queries they ran, what they exported and when. Check whether logs can be altered or deleted by a user.
Why it matters: Chain of custody now extends into the analysis environment. Courts have started asking.
11. Court output and expert support
Test: Produce a complete examination report from a real case. Assess whether it is comprehensible to a non-technical reader, whether it states methodology and limits, and whether exhibits can be regenerated from the log months later.
Why it matters: The report is the deliverable. Everything before it is preparation.
Scoring your forensic video software pilot
Weight these unevenly. Items 3, 4, 5, 9 and 10 are pass or fail – a product that fails any of them creates a structural risk in every case it touches, regardless of how good its analysis is. Items 1, 6, 7, 8 and 11 are graded, and reasonable agencies will weight them differently depending on caseload profile. Item 2 sits in between.
Two further questions belong in the commercial evaluation rather than the technical one. What is the total cost at your actual concurrency, including per-seat licensing, retraining and support? And what happens to your case data and your processing logs if you stop paying – can they be exported in an open format, or are they hostage to the platform?
A note on demonstrations
Insist that any demonstration uses footage you supply, seized in the way your officers actually seize it. The gap between vendor sample footage and real Indian CCTV – low bitrate, aggressive compression, mixed lighting, dusty lenses, drifting clocks – is where most disappointment originates. Platforms like pi-sense are designed to be evaluated exactly this way.
