More in AI & ML — page 10
PTQ versus QAT for model quantization
WHAT IT TESTS: quantization workflows. OUTLINE: PTQ quantizes a trained model with calibration data, fast but may lose accuracy; QAT simulates quantization in training so weights adapt. RED FLAG: thinking QAT needs no data or PTQ is always lossless.
Why averaging shots reduces image noise
WHAT IT TESTS: noise statistics in imaging. OUTLINE: random noise averages out while signal stays, so SNR rises with the square root of frame count; limits include motion and fixed-pattern noise. RED FLAG: claiming noise drops linearly with frames.
The role of anchor boxes in detectors
WHAT IT TESTS: anchor-based detection. OUTLINE: anchors are reference boxes of set scales and ratios; the model regresses offsets, framing detection as classification plus refinement; downsides are tuning, imbalance. RED FLAG: anchors are not final boxes.
Triangulating 3D points from two views
WHAT IT TESTS: 3D reconstruction geometry. OUTLINE: back-project each 2D point into a ray, intersect rays from both cameras using known poses, solve via DLT and refine by minimizing reprojection error. RED FLAG: ignoring noise so rays never exactly meet.
Single vs multi-image super-resolution
WHAT IT TESTS: super-resolution methods and losses. OUTLINE: SISR hallucinates detail from one image, multi-image fuses sub-pixel shifts; pixel losses blur, perceptual or GAN losses add realism. RED FLAG: thinking L2 loss gives the sharpest result.
Non-maximum suppression in detection
WHAT IT TESTS: detector post-processing. OUTLINE: detectors emit many overlapping boxes per object; NMS keeps the highest-scoring box and removes others above an IoU threshold. RED FLAG: forgetting NMS is per-class or misstating why duplicates arise.
One-stage vs two-stage detectors
WHAT IT TESTS: detector architecture trade-offs. OUTLINE: one-stage predicts boxes directly for speed; two-stage proposes then refines for accuracy; focal loss narrows the gap. RED FLAG: claiming one is always better without the speed-accuracy framing.
Designing a multi-object tracker
WHAT IT TESTS: tracking-by-detection design. OUTLINE: detect per frame, predict motion with a filter, associate via IoU and appearance, manage track lifecycle, handle occlusion with re-ID. RED FLAG: ignoring data association or ID switches.
Challenges deploying a model on edge hardware
WHAT IT TESTS: edge deployment constraints. OUTLINE: limited memory and compute cause latency, thermal and power limits, accuracy loss from compression, operator support gaps. RED FLAG: only mentioning accuracy while ignoring latency, memory, and power.
When a homography is a valid model
WHAT IT TESTS: projective geometry assumptions. OUTLINE: homography holds for pure rotation or a planar scene; it fails with translation plus 3D parallax, where epipolar geometry applies. RED FLAG: thinking homography works for any two views.
Contrastive learning vs masked image modeling
WHAT IT TESTS: self-supervised pretraining. OUTLINE: contrastive aligns augmented views via instance discrimination; MAE reconstructs masked patches; they differ in augmentation and fine-tuning. RED FLAG: conflating them or saying one always wins.
Zero-shot classification with CLIP
WHAT IT TESTS: contrastive vision-language inference. OUTLINE: encode image and label prompts into a shared space, compare via cosine similarity, pick the highest. RED FLAG: thinking CLIP needs fine-tuning or a fixed label head to classify.
Point cloud vs voxel grid vs NeRF
WHAT IT TESTS: choosing 3D scene representations. OUTLINE: point clouds are sparse and fast but unstructured, voxels are regular for collision checks but memory-heavy, NeRFs render photorealistically but are slow. RED FLAG: declaring one universally best.
Cross-attention in transformer VQA models
WHAT IT TESTS: vision-language fusion via attention. OUTLINE: text queries attend over image regions, learning alignment that grounds words to visual content. RED FLAG: confusing self-attention with cross-attention or saying it just concatenates features.
The data association problem in SLAM
WHAT IT TESTS: correspondence reliability in SLAM. OUTLINE: matching observations to landmarks, why wrong matches corrupt the map, robust techniques like RANSAC and descriptor matching. RED FLAG: treating it as trivial or ignoring outlier rejection.
Core components of visual SLAM
WHAT IT TESTS: SLAM system structure. OUTLINE: tracking estimates per-frame pose, mapping builds and refines the 3D map, loop closure detects revisits and corrects drift. RED FLAG: confusing tracking with mapping or omitting loop closure.
Adapting a 2D CNN for video action recognition
WHAT IT TESTS: extending image models to temporal data. OUTLINE: run the 2D CNN per frame, pool features over time, optionally add two-stream or 3D conv. RED FLAG: assuming a single frame suffices or ignoring temporal information entirely.
Deploying segmentation on edge devices
WHAT IT TESTS: model compression for edge. OUTLINE: pick efficient architectures, apply INT8 quantization, distill from a large teacher. RED FLAG: only shrinking input resolution or naming one technique without the accuracy-versus-latency trade-off.
Improving small object detection
WHAT IT TESTS: practical detector debugging. OUTLINE: raise input resolution and tile, use feature pyramids for high-res features, tune anchors and copy-paste augmentation. RED FLAG: only saying collect more data without addressing scale.
What is Bundle Adjustment and why is it tractable?
WHAT IT TESTS: structure-from-motion optimization. OUTLINE: jointly refine 3D points and camera poses by minimizing reprojection error, expensive due to many coupled parameters; sparsity of the Jacobian and the Schur complement make it tractable.