src/modules/tensor/cpu/kernel/random_erase.cpp:112

• Noise indexing here uses hard-coded 255 values instead of RANDOM_ERASE_NOISE_BUFFER_SIDE, making the code inconsistent with the rest of the file and fragile if the constant ever changes. Use the macro consistently.

                Rpp32u noiseRowOffset = ((y1 + i + batchCount) % 255) * 255 * 3;
                for (int j = 0; j < boxWidth; j++)
                {
                    Rpp32u noiseIdx = noiseRowOffset + ((x1 + j) % 255 * 3);

api/rppt_tensor_effects_augmentations.h:510

• The Doxygen image reference uses effects_augmentations_random_erase_dropout_img150x150.png, but the added file path is docs/data/doxygenOutputs/effects_augmentation_random_erase_dropout_img150x150.png (missing the s in augmentations). This will break the generated docs image. Rename the file or update the \image html reference to match.

 * \image html img150x150.png Sample Input
 * \image html effects_augmentations_random_erase_dropout_img150x150.png Sample Output
 * \param [in] srcPtr source tensor in HIP memory (for HIP backend) or HOST memory (for HOST backend)

utilities/test_suite/rpp_test_suite_image.h:1747

• init_dropout_erase() forces x_start/y_start to be at least 1.0 even when x_slack/y_slack is 0 (box equals ROI size). This can push the box outside the ROI or make the erase region empty for small ROIs. Allow 0 start (and clamp to slack) so boxes can start at the ROI origin.

        const float x_slack = std::max(0.0f, roiW - boxW);
        const float y_slack = std::max(0.0f, roiH - boxH);

        const float x_start = std::max(1.0f, std::min(pos_ratio(rng) * x_slack, x_slack));
        const float y_start = std::max(1.0f, std::min(pos_ratio(rng) * y_slack, y_slack));

utilities/test_suite/rpp_test_suite_image.h:1753

• RpptRoiLtrb uses inclusive right/bottom coordinates (see compute_xywh_from_ltrb_host() which does rb - lt + 1). Here rb is set to lt + boxW/boxH, which makes the box 1 pixel larger than intended and can overshoot the ROI. Consider setting rb to lt + boxW/boxH - 1 (after clamping).

        RpptRoiLtrb &box = anchorBoxInfoTensor[i * maxBoxesPerImage];
        box.lt.x = static_cast<Rpp32u>(roiX + x_start);
        box.lt.y = static_cast<Rpp32u>(roiY + y_start);
        box.rb.x = static_cast<Rpp32u>(roiX + x_start + boxW);
        box.rb.y = static_cast<Rpp32u>(roiY + y_start + boxH);