Pranay Obla Anandbabu

Add pose estimation on top of object detection over tune the detector harder

Concealment is defined by what the body does, not by what's in the frame. A bounding box around a person and a product can't distinguish picking an item up from hiding it. Keypoint dynamics can.

Trade-off · A second model in the pipeline, which costs inference time on every frame that contains a person.

Sequence models (YOLO-LSTM, RCNN-LSTM) over per-frame classification

Shoplifting is an action, and an action only exists across frames. Per-frame scoring fires on any single ambiguous posture, which is exactly where the false positives were coming from.

Trade-off · Needs temporally labelled data and a buffer of frames before it can decide, so alerts arrive later.

Build a custom dataset from public traffic footage over an off-the-shelf traffic dataset

Lane violations are defined by local road markings and driving conventions. A model trained on someone else's roads learns the wrong geometry.

Trade-off · Weeks of scraping and annotation before a single model could be trained.

Benchmark five architectures before committing over start with the obvious YOLOv7 baseline

Real-time was a hard requirement, so the choice had to be made on the accuracy-versus-latency curve for this data, not on published numbers from a different benchmark. YOLO-NAS won and shipped.

Trade-off · Spent early project time on evaluation instead of building.

Streamlit and Vercel for the frontend over a custom web application

The deliverable was a working demonstration on a two-month clock. The interface needed to exist and be shareable, not to be bespoke.

Trade-off · Not the deployment you'd choose for real production traffic.

Post-training quantization over quantization-aware training

QAT means retraining, and the detection models had already been trained on millions of labelled images. Redoing all of that to recover a few points of post-quantization accuracy was not a trade worth making.

Trade-off · PTQ gives up more accuracy than QAT, so the retention-versus-latency curve had to be measured and documented rather than assumed.

Port YOLO-NAS's quantization block into YOLOv7 over migrate production to YOLO-NAS

YOLO-NAS ships a quantization-friendly block that YOLOv7 doesn't have, but production was already running on v7. Transplanting the block keeps the existing weights and pipeline intact; switching architectures would have meant retraining and revalidating the whole detection stack.

Trade-off · Implementing a block on an architecture it wasn't designed for means more integration work, with no guarantee the quantization gains carry over.